Z Zhang (@1.75) vs J Nam (@2.1)

Our Prediction:

Z Zhang will win

Z Zhang – J Nam Match Prediction | 15-09-2019 01:00

Funding: This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (NRF-2018M3A9A7053266), the Bio-Synergy Research Project (NRF-2017M3A9C4092978) of the Ministry of Science and ICT through the National Research Foundation. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

With a convolutional neural network (CNN), the authors extracted local residue patterns to predict the binding affinity between drugs and targets. As a result, their model exhibited better performance on a kinase family bioassay dataset [25, 26] than the previous model, kronRLS [27] and SimBoost [28]. In a paper by ztrk et al., DeepDTA was used to represent raw sequences and SMILES as one-hot vectors or labels [24]. Because their model is optimized by densely constructed kinase affinities, DeepDTA is appropriate to predict kinase affinities not to predict new DTIs with various protein classes. Furthermore, they evaluated their performances on the identical dataset, rather than on independent dataset from new sources or databases. One way to reduce the loss of feature information is to process raw sequences and SMILES as their forms.


Although we cannot measure exactly how those values affect the DTI prediction results, the pooled maximum convolution result will affect the prediction performance by going through higher fully connected layers. Therefore, if our model is capable of capturing local residue patterns, it would give high values to important protein regions, such as actual binding sites. Because we pooled the maximum convolution results by each filter for each window, the pooled results could highlight regions of matches with local residue patterns.

Bet365 could be streaming this event live. On Extratips.com you can watch the M Gong/Z Zhang vs J Nam/M Song match that starts at 05:00 on 15 September 2019. If you are a registered member you can watch M Gong/Z Zhang vs J Nam/M Song video highlights are welcome from visitors in case the live broadcast link is broken. Please check what games is bet365 streaming by using the player above. We can not be held responsible for third party video content so please forward any claims to video file owners.

From the queried binding sites and pooled maximum convolution results, we statistically test our assumption that the pooled maximum convolution results cover the important regions, including binding sites. The sc-PDB database provides atom-level descriptions of proteins, ligands, and binding sites from complex structures [40]. We visualized two high-score sc-PDB entries from two perspectivesthe whole receptor-ligand complex and binding site-ligand perspectivesby using UCSF Chimera [42] as shown in Fig 6. For each normal distribution constructed by the randomly generated convolution results, considered a null hypothesis, we executed a right-tailed t-test with the number from the convolution results of our model for each window. The sc-PDB entry information and p-values of a window for each sc-PDB entry are summarized in the S1 File. We visualized two sc-PDB entries, 1a7x_1 and 1ny3_1. Second, we rendered residues covered by convolution results by the number of covering convolution results. 1ny3_1 is the complex of the kinase protein, MAP kinase-activated protein kinase 2 (MAPK2_HUMAN in UniProt) with sequence length 400, and ADENOSINE-5-DIPHOSPHATE (ADP in PDB ligand) [44]. To visualize convolution results with a simplified view, first, we selected the top 5 ranked globally max-pooled results among all filters for each window because whole protein sequences are usually covered by convolution results if we select all results. Thus, we randomly generated 128 convolution results 10,000 times for each sc-PDB entry and counted how many of those random results covered each amino acid in the binding sites, which resulted in the construction of normal distributions. Because we did not know which window detects the binding site, we took the most significant p-value (minimum p-value adjusted by the Benjamini-Hochberg procedure [41]). Examining and validating the convolution results from the intermediate layer showed that our model could capture local residue patterns that participate in DTIs. In addition, we examined sc-PDB entries with the most significant p-values for diverse window sizes. Each window has 128 pooled convolution results, which shows bias in covering some regions. 1a7x_1, representing the complex of the ion channel, protein Peptidyl-prolyl cis-trans isomerase FKBP1A (FKB1A_HUMAN in UniProt), which has a short sequence length (108), and BENZYL-CARBAMIC ACID [8-DEETHYL-ASCOMYCIN-8-YL]ETHYL ESTER (FKA in PDB ligand) [43]. Through the above evaluation, we can confirm that our proposed model is capable of capturing local residue patterns of proteins that are considered important features for DTI prediction, such as actual binding sites. We summarize the results of binding site detection from the most significant p-value among windows by significance level cutoff in Fig 5. By parsing binding site annotations, we can query binding sites between protein domains and pharmacological ligands for 7,179 entries of Vertebrata.


Docking methods recruit various scoring functions and mode definitions to minimize free energy for binding. To overcome the requirement of the bipartite model for massive computational power, Beakley et al. developed the bipartite local model, which trains the interaction model locally but not globally. As another approach to DTI prediction models, matrix factorization methods have been recruited to predict DTIs, which approximate multiplying two latent matrices representing the compound and target protein to an interaction matrix and similarity score matrix [5, 6]. Among computational approaches, docking methods, which simulate the binding of a small molecule and a protein using 3D structure, were initially studied. In addition, studies have examined several similarity-based methods in which it was assumed that drugs bind to proteins similar to known targets and vice versa. However, similarity-based methods are not commonly used at present to predict DTIs, as researchers have found that similarity-based methods work well for DTIs within specific protein classes but not for other classes [7]. In addition, some proteins do not show strong sequence similarity with proteins sharing an identical interacting compound [8]. Docking methods have advanced by themselves, and recently, the Docking Approach using Ray-Casting (DARC) model identified 21 compounds by using an elaborate binding pocket topography mapping methodology, and the results were reproduced in a biochemical assay [2]. In addition to substantially reducing the computational complexity, this model exhibited higher performance than the previous model [4]. One of the early methods is that of Yamanashi et al., which utilized a kernel regression method to use the information on known drug interactions as the input to identify new DTIs, combining a chemical space and genomic spaces into a pharmacological space [3]. In this work, regularized matrix factorization methods successfully learn the manifold lying under DTIs, giving the highest performance among previous DTI prediction methods.

As a result, the detected local features of protein sequences perform better than other protein descriptors for DTI prediction and previous models for predicting PubChem independent test datasets. In this work, we constructed a novel DTI prediction model to extract local residue patterns of target protein sequences using a CNN-based deep learning approach. However, identifying drug candidates via biological assays is very time and cost consuming, which introduces the need for a computational prediction approach for the identification of DTIs. That is, our approach of capturing local residue patterns with CNN successfully enriches protein features from a raw sequence. Therefore, identification of DTIs is a crucial step in drug discovery. Drugs work by interacting with target proteins to activate or inhibit a targets biological process.

After processing both the drug and protein layers, we concatenated these layers and constructed the fully connected layer, resulting in the output. We extracted the local residue patterns from protein sequences via CNN and yielded a latent representation of drug fingerprints via fully connected layers. Every layer except the output layer was activated with the exponential linear unit (ELU) function [51].

Zhang Z. Nam Ji S. PREDICTION (15.09.2019)

Finally, each drug can be represented as a binary vector with a length of 2,048, whose indices indicate the existence of specific substructures. In our model, we used the raw protein sequence as the input for the protein but did not use the raw SMILES string as the input for the drug. For the drug, we used the Morgan/Circular drug fingerprint, which analyzes molecules as a graph and retrieves substructures of molecular structures from subgraphs of the whole molecular graph [21]. Specifically, we used RDKit [50] to yield a Morgan/Circular fingerprint with a radius of 2 from a raw SMILES string.

It's also easy to find video highlights and news from the most popular sports leagues in the world. on your iPhone, iPad, Android or Windows phone. Nam J. If this match is covered by bet365 live streaming you can watch tennis match Zhang Z. live stream online if you are registered member of bet365, the leading online betting company that has streaming coverage for more than 140.000 live sports events with live betting during the year. We don't offer a TV schedule here, if you would like to watch this match on TV you'll probably find it it on some more popular channels like iTV, BBC, Al Jazeera Sports, Sky Sports, Gol TV, Canal+, SportTV, FOX Soccer, Setanta, ESPN, etc. You can watch Zhang Z. Live bet365 odds are viewable in SofaScore's tennis livescore section. Nam J. vs.

We measured various performances such as sensitivity (Sen.), specificity (Spe.), precision (Pre.), accuracy (Acc.), and F1 score (F1) from the prediction results given by descriptors (A-D). (B) PubChem dataset whose compounds are not in the training dataset. (C) PubChem dataset whose targets are not in the training dataset. (A) All queried PubChem datasets. (D) PubChem dataset whose compounds and targets are not in the training dataset. Our convolution model shows better performances for all datasets in terms of accuracy and F1 score.


Therefore, we independently compared performances between DL-CPI and our model by additionally built the training, validation, and test datasets. Because protein descriptor of DL-CPI is sparse, containing few values in large dimension, which may decrease performances. We confirmed that the proposed model shows better performance than DL-CPI. which used protein domain information. In addition to the three models we compared, we also compared our model with DL-CPI [23] built by Tian et al. For proteins whose domain information is not in Pfam [39], datasets for training, validation and test are not fully available. Performance comparison results are described in Fig E in S1 Text.

Second, DeepDTI built by Wen et al. In addition to the comparison between convolution in our model and other protein descriptors, in this section, we compared the performance of our model against recently developed deep-learning-based models. We tested the validity of implemented MFDR and confirmed that the implemented model produces reasonably same performance compared to the results from its original work (see Fig C S1 Text). We also used DeepDTA with the code from the original work (https://github.com/hkmztrk/DeepDTA) and optimized hyperparameters they provided. For the DTI prediction performance comparison, we activate the last layer with sigmoid function to predict interaction, not affinity, also we changed loss function as binary cross-entropy from mean squared error. As a result, their model gives better performances than previous bipartite local models. DeepDTI takes amino acid, dipeptide and tripeptide compositions (protein sequence composition descriptors, PSC) as the protein input and ECFP with radius 1, 2 and 3 as the compound input. Third, DeepDTA built by Ozturk et al. DeepDTA is optimized for Davis [25] and KIBA [26] dataset which contains kinases protein, their inhibitors, and dense affinity values, showing better prediction performances than previous affinity prediction models. is based on DBN [20], which is a stack of restricted Boltzmann machine (RBM). We selected three deep learning models for comparisons, SAE (MFDR, Peng et al, 2016) [22], DBN (DeepDTI, Wen et al, 2017) [7] and CNN (DeepDTA, Ozturk et al, 2018). First, MFDR trains SAE in an unsupervised manner, while proteins are represented by multi-scale local descriptor feature [38] and compounds are represented by PubChem fingerprints as input and output for SAE. It should be noticed that we compared the performance of all three models by training and testing with the same data set we used for a fair comparison. used stacked CNN on protein sequences and SMILES to predict affinity between target protein and compound. We used DeepDTI with the code that the authors provided (https://github.com/Bjoux2/DeepDTIs_DBN) and optimized hyperparameters as the authors mentioned. With trained deep representations of sparse Auto-Encoder, they performed 5-fold cross-validation by using SVM. Because the authors do not provide the model, we implemented the MFDR model with optimized parameters the author provided in their original paper.