Study On The Structure-Activity Relationship Of JAK1 Inhibitors

As a member of the Janus kinase family,Janus kinase 1（JAK1）plays an important role in immune response,hematopoiesis,cell growth and proliferation.It is one of the important targets of many inflammatory and immune diseases.As more and more small molecule drugs are discovered,the importance of computational chemistry and chemoinformatics in the drug design has been particularly prominent.Chemoinformatics uses computer simulations to process massive amounts of compound data to mine potential chemical information to guide drug discovery.Here,various forms of molecular characterization methods and machine learning methods are used to establish classification models and regression models for JAK1,as well as JAK1 inhibitor/JAK2 inhibitor selectivity models.Finally,important structures are analyzed.The main research contents are as follows:（1）Study on the classification model of high/weak activity of JAK1 inhibitorsIn this part of the experiment,2982 JAK1 inhibitors were collected.The inhibitors were divided into highly and weakly active with the threshold of IC₅₀=50 n M,100 n M,respectively.The dataset was divided using random methods（training set:test set=2236:746）and self-organizing neural mapping method（SOM,training set:test set=2219:763）.The molecules were characterized using MACCS and ECFPs,using support vector machines（SVM）,Decision Trees（DT）,Random Forests（RF）,Extreme Gradient Boosting（XGBoost）,Deep Neural Networks（DNN）,and Convolutional Neural Networks（CNN）to built 22 classification models.Among them,the accuracy rate（ACC）and Matthews correlation coefficient（MCC）of the best model Model 5B on the test set reached 94.5%and 0.89,respectively.Based on the dendrogram,the fingerprints of MACCS and ECFPs were analyzed.The important structural fragments that may affect the bioactivities of JAK1 inhibitors were summarized.In addition,10 selective models were established based on 904JAK1/JAK2 selective inhibitors.The accuracy rate of the best model Model 11B on the test set reached 85.5%,and the Kappa coefficient and Jaccard coefficient both reached 0.75.（2）Regression prediction model of JAK1 inhibitorsIn this part,537 Jak1 inhibitors were collected,their bioactivities were measured by mobility shift assay,and the data set was randomly divided four times.Their CORINA descriptors and RDkit descriptors were computed.Using the grid optimization method based on five-fold cross-validation to optimize the parameters of SVM and RF algorithms.A total of 64 regression models in 16 groups were established.The R² and MSE of the best three models Model 13C,Model 15C and Model 18C on the test set were 0.6418,0.6798,0.6782 and 0.2505,0.2239,0.2250,respectively.Descriptor analysis showed that descriptors describing the properties of molecularπatomic electronegativity,effective atomic polarizability,πatomic charge,σatomic charge,lone-pair electron electronegativity,and total atomic charge appeared simultaneously in the three best models.These properties are closely related to the interaction between the cyano group in the inhibitor and the JAK1 receptor,and basically consistent with the interaction in the JAK1 crystal structure（3EYG）In this paper,22 well-performed classification models on JAK1inhibitors,10 selective classification models of JAK1/JAK2 inhibitors,and64 regression models on JAK1 inhibitors were constructed using SVM,DT,RF,XGBoost,DNN and CNN algorithms.In addition,this paper summarized the structural fragments that are strongly related to the activity of JAK1 inhibitors,and analyzes the interaction between the inhibitor and the receptor,which has guiding significance for the discovery and design of new JAK1 inhibitor drugs.