Computer Aided QSAR Study Of Cyclooxygenase 1 And 2 Inhibitors Activity

Human cyclooxygenase?COX? family is composed of three isoenzymes: COX-1 is similar with COX-2, and COX-3 is also an isoenzyme. COX-3 is a variant of COX-1, mainly existing in the brain cortex and heart, which is not explored in detail here; COX-1 ande COX-2 were found in earlier time and they can catalyze the formation of prostaglandins, and thromboxane. The prostaglandins are autocoid mediators that affect virtually all known physiological and pathological processes via their reversible interaction with G-protein coupled membrane receptors. COX enzymes are clinically important because they are inhibited by aspirin and numerous non-steroidal anti-inflammatory drugs ?NSAIDs?. This inhibition of COX confers relief from inflammatory, pyretic, thrombotic, neurodegenerative and oncological maladies. In this study, COX-1 and COX-2 inhibitors were collected as many as possible, and support vector machine method?SVM? and multiply linear regression method ?MLR? were used to build uantitative prediction models, as same as Self-Organizing Map ?SOM? method, in order to explore COX inhibitors structure-activity relationship. The content of this study included two parts, described as follows:1. COX-1 inhibitors were collected for building database; and qualitative classification models of active and weakly active inhibitors of COX-1,as well as quantitative prediction models,were established.?1? Qualitative classification models of active and weakly active COX-1 inhibitors:1862 COX-1 inhibitors, 10?M as the threshold of active inhibitors and weak active inhibitors, were split by Self-Organizing Map ?SOM? method into training set with 1310 compounds and test set with 552 compounds. Use ADRIANA.Code software to calculate descriptors of COX-1 inhibitors. Then, methods, Pearson correlation analysis was used to screen descriptors as same as stepwise regression method.166 MACCS-Fingerprint descriptors were also prepared for model building. SOM and SVM methods were used to construct classification models. The obtained model all showed a good predictive accuracy of over 85% on the test set. Model2D was the optimum classification model in this work. The classification accuracy rate of test set was 89.02%. The AUC of test set was 0.86, and five-fold/ten-fold cross-verification and Y-randomization validation showed credibility of the models. Extended Connectivity Fingerprints ?ECFP₄? was conducted to explore the relation between COX-1 inhibitors space structure and bioactivity.?2? Quantitative prediction models of COX-1 inhibitors:Multiple Linear Regression and SVM methods were used to build quantitative prediction models of COX-1 inhibitors.357 COX-1 inhibitors were divided by Self-Organizing Map ?SOM? method into training set with 220 compounds and test set with 137 compounds. Similarly, the structure parameters of ADRIANA.Code for each molecule were calculated, and the best structure parameters set was selected by using correlation analysis. To screen COX filter descriptors, stepwise regressions were used at the same time. Multiple Linear Regression and SVM methods were used to construct quantitative models. The models of training set which were generated by SOM method were better than that were generated by random method. For modeling methods, SVM method was more effective than MLR method. The robust model showed multiple correlation coefficient ?R? of training set up to 0.88, root-mean-square error ?RMSE? was 0.30. The test set showed R=0.78, RMSE=0.26. And the best model was robustness, verified by Y-randomization method.2. COX-2 inhibitors were collected for building database; and qualitative classification models of active and weakly active inhibitors of COX-2,as well as quantitative prediction models,were established.?1? Qualitative classification models of active and weakly active COX-2 inhibitors:2717 COX-2 inhibitors, 10?M as the threshold of active inhibitors and weak active inhibitors, were split into training set with 1971 compounds and test set with 746 compounds by SOM method. Use ADRIANA.Code software to calculate descriptors of COX-2 inhibitors. Then, Pearson correlation analysis was used to screen descriptors, as same as stepwise regression method.166 MACCS-Fingerprint descriptors were also prepared for model building. SOM and SVM methods were used to construct classification models. The obtained models all showed good predictive accuracies of over 80% on the test set. Model2D was the optimum classification model in this work. The classification accuracy rate of test set was 89.02%. The AUC of test set was 0.86, and five-fold/ten-fold cross-verification and Y-randomization validation showed credibility of the models. We calculated the ECFP₄ descriptors COX-2 inhibitors, and analyzed segments of substructure which contributed more effictively to highly active inhibitors in order to explore the relation between COX-2 inhibitors space structure and bioactivity.?2? Quantitative prediction models of COX-2 inhibitors:575 COX-2 inhibitors were split by SOM method into training set with 372 compounds and test set with 203 compounds. Similarly, the structure parameters of ADRIANA.Code for each molecule were calculated, and the best structure parameters set was selected by using correlation analysis. To screen COX filter descriptors, stepwise regressions were used at the same time. Then the prediction models were established based on the selected descriptors. For all the prediction models, the R on training set were higher than 0.80, RMSE were less than 0.35; R on test sets were higher than 0.75, RMSE were less than 0. 35. Models were robustness, verified by Y-randomization method.In this study, the classification models and quantitative prediction models, which we built, performed a good forecasting ability, in order to provide guidance for design and research new anti-inflammatory and anti-tumor medicine, etc.