GS-SVM Based Prediction Model Of Telomerease And ATR Inhibitors

Ataxia Telangiectasia and Rad3-related(ATR)and Telomerase are two important anti-tumor targets that can cause tumor cell death but not affect normal cells.Although a few ATR and Telomerase inhibitors were developed,there is still lack to develop new drug.In this study,we tried to construct high performance computer prediction model to predict new Telomerase and ATR inhibitors.This study mainly was divided into the following five parts:Part ?.ReviewIn this section,the information focused on the advances in anti-tumor of ATR and Telomerase inhibitors,and the progress of machine learning methods in drug discovery was reviewed.Firstly,the function of ATR and Telomerase,as well as the role of ATR and Telomerase in telomere elongation were introduced.Secondly,the information such as structure,activity regarding inhibitors of ATR and TEL was summarized.Finally,the application of machine learning method in drug discovery and its corresponding advantages and disadvantages were introduced.This section provided both theoretical reference and data basis for caring out the study.Part ?.Data collection and processingIn this section,the ATR and Telomerase inhibitors and inactive compounds were collected from the international authoritative database BindingDB and ZINC respectively,and randomly divided into training set and test set.As results,ATR training set consisting of 292 compounds,ATR test set consisting of 145 compounds,Telomerase training set consisting of 301 compounds,and Telomerase test set consisting of 150 compounds were obtained(each set includes inhibitor and inactive compounds).Then,the molecular descriptors of the training sets were calculated and 94 descriptors for ATR,92 descriptors for Telomerase were selected based on the relationship between the descriptors and biological activity of the compounds.Finally,the descriptors were further analyzed by PCA,and 22 components of the ATR descriptors and 16 components of the Telomerase inhibitor descriptors were selected for further model construction.Part ? .Construction of ATR and Telomerase inhibitor SVM modelsIn this section,the parameters include cross validation mode,SVM type and kernel function of SVM model were determined by using training set via the single variable principle.When set cross validation mode to three-fold cross validation,SVM type as c-SVC and the kernel function as radial basis,the ATR inhibitor SVM models(SVM-ATR)and the Telomerase inhibitor SVM models(SVM-TEL)achieved the highest accuracy.Finally,The SVM-ATR and SVM-TEL models were validated and evaluated by using their test sets respectively.The results show that the sensitivity of the SVM-ATR model was 100%,the specificity was 95.74%,the prediction accuracy was 97.24% and the Matthews correlation coefficient was 94.22%,while the sensitivity,specificity,prediction accuracy and the Matthews correlation coefficient of the SVM-TEL model was 70%,97.77%,88.67%,74.36% respectively.It indicated that the SVM-ATR and SVM-TEL models have high predictive performance.Part ?.Optimization of SVM-ATR and SVM-TEL modelsIn this section,in order to improve the performance of SVM-ATR and SVM-TEL model,the parameter optimization algorithm(grid search algorithm,particle swarm optimization and genetic algorithm)were used to optimize the penalty parameters c and kernel function parameters g for SVM-ATR and SVM-TEL models.The optimal SVM-ATR model and SVM-TEL model constructed with optimal parameters were then validated and evaluated by using the test set.The results showed that both SVM-ATR model optimized using grid search algorithm(GS-SVM-ATR)and SVM-TEL model optimized using grid search algorithm(GS-SVM-TEL)achieved the best performance.When the penalty parameter c sets to 0.35255 and the kernel function parameter g sets to 0.25,the accuracy of cross validation of GS-SVM-ATR model was 93.1507%,the sensitivity was 100%,the specificity index was 100%,the prediction accuracy was 100%,and the Matthews correlation coefficient was 100%.For GS-SVM-ATR model,when the penalty parameter c set to 1.4142 and the kernel function parameter g set to 0.7071,the accuracy of cross validation of GS-SVM-ATR model was 86.0465%,the sensitivity was 88%,the specificity index was 91%,the prediction accuracy rate was 90%,and the Matthews correlation coefficient was 77.91%.Part?.Prediction of new ATR and Telomerase inhibitorsIn this section,a novel set consisting of 59 compounds(activity unknown)was predicted using the optimal ATR and Telomerase inhibitor prediction models(GS-SVM-ATR and GS-SVM-TEL model)and verified by experiment to validate the prediction performance in practice.As the results showed,the GS-SVM-TEL model obtained 18 potential telomerase inhibitors from the new compounds,and four available compounds among 18 compounds were tested using MTT.Finally,three compounds showed strong activity of tumor inhibition,and Molecular docking also revealed that all of three compounds can bind to active pocket of telomerase.However,due to the limitation of the number of compounds,the GS-SVM-ATR model did not obtain ATR inhibitors from the new set.In order to verify the performance of the GS-SVM-ATR model,23 ATR inhibitors reported in the literature were collected as additional validation set,and 22 compounds of them were predicted to be positive and the accuracy rate was 95%.Therefore,the GS-SVM-ATR model and the GS-SVM-TEL model constructed in this study have high accuracy and reliability for the prediction of inhibitors.In summary,the high performance prediction model both for Telomerase and ATR inhibitor prediction were successfully constructed,these models can provide useful support for further development of Telomerase and ATR inhibitors,and provide reference for development of other types of drug.