Computer Modeling Of Structure-Activity Relationship In Human 5-Lipoxygenase Inhibitors

Inflammation is a physical process caused by infection, injury and stimulations, which shows redness, swelling, fever and pain in clinic. It may damage our health and cause disorder in body.5-lipoxygenase （5-LOX） is a key enzyme in arachidonic acid metabolism network, which is responsible for the regulation of inflammatory process. The study of human 5-LOX inhibitors will contribute to design effective drugs treating diseases like cancer, asthma and arthritis. In our research, we used computer-aided drug design to build structure-activity relationships of human 5-LOX inhibitors and conduct virtual screening in order to design new drugs.Methods including self-organizing neural network （SOM）, support vector machine （SVM）, Naive Bayes, Radial Basis Function network and Multilayer Perceptron network were used to build 16 classification models.220 compounds were split by SOM into training set with 145 compounds and test set with 75 compounds. Threshold for distinguishing high and weak activity inhibitors was 10μM. We calculated descriptors with ADRIANA.Code and MACCS descriptors, using enter regression, stepwise regression, rank by SVM weight and F-score to select them. The best model consists of 14 descriptors selected by SVM weight. The prediction results of training set by SOM model, Model3A, were 83.22%（Q）,0.65 （MCC）, and those of test set were 86.49%（Q）,0.73（MCC）. The prediction results of training set by SVM model, Model3B, were 80.00%（Q）,0.58（MCC）, and those of test set were 82.67%（Q）,0.64 （MCC）. Extended-Connectivity Fingerprints （ECFP₄） analysis suggested the importance of hydrogen bond, number of single rotatable bonds and hydrophobic.Multiple Linear Regression （MLR） and Support Vector Machine （SVM） algorithms were used to construct two quantitative models.180 compounds tested by fluorescent signal were split by SOM into training set with 120 compounds and test set with 60 compounds. After calculating descriptors, we selected 12 descriptors based on T test value in enter regression. The value of correlation coefficient of training set predicted by ModelMLR was 0.78 and that of test set was 0.72, while the value of training set predicted by ModelQSVM3 was 0.85 and that of test set was 0.74. The results showed that hydrogen bond, electrostatic property and stereoscopic chemistry were related to activity.We got 22 compounds from commercial databases by virtual screening based on shape, color and electrostatic similarity. Models including Model3A, Model3B, ModelNB1, ModelRBF1 and ModelMLPl were used to make further prediction. Then we docked 19 compounds with high activity in prediction to the crystal structure of human 5-LOX. Docking analysis showed that there were some interactions including π-π stacking interactions, complementary electrostatic potential and hydrogen bonds.