Kinome-Wide Polypharmacology Profiling Of Small Molecules By Multi-Task Graph Network Approach

Protein kinases are one of the largest protein families,consisting of more than 500 members,accounting for 1.7% of the human genome.Protein kinases can catalyze the transfer of the terminal phosphate group on adenosine triphosphate(ATP)to a specific substrate and play a pivotal role in cell signaling.Dysregulation of protein kinases is closely related to the occurrence and development of cancer,central nervous system diseases,inflammatory diseases,cardiovascular diseases and other diseases.A series of small-molecule inhibitors targeting protein kinases have been developed for the treatment of related diseases.Accurate detection of the interactions between small molecules and protein kinases is quite critical to various related applications in drug development,such as discovery of lead compounds,drug repurposing,and elucidation of potential off-target effects.At present,some ligand-based machine learning methods have been employed to predict the interactions between small molecules and protein kinases.However,most of these studies used traditional machine learning models based on molecular fingerprints/descriptors,whose prediction performance is highly dependent on the selection of molecular fingerprints/descriptors.In addition,most of these predictive models lack interpretability,which hinders their application in drug design.In this thesis,we used Graph Neural Network(GNN)algorithm to develop a model to predict the inhibitory activity of small molecules against different protein kinases and developed a model-agnostic interpretable method to provide interpretable analysis for the predictive results of the GNN-based model.The main contents and results of this study are as follows:(1)Firstly,nearly 260000 data from different public databases were collected,and a predictive model called Auxiliary Multi-task Graph Isomorphism Network with Uncertainty weighting(AMGU)was developed.The model can simultaneously predict the inhibitory activity of small molecules against 204 different kinases.The results show that the AMGU model has better performance on the internal test set than the descriptorbased model and the most advanced GNN models.Furthermore,it also exhibited much better performance on two external test sets,indicating that the AMGU model has enhanced generalizability due to its strong transfer learning capacity.(2)Then,a na(?)ve model-agnostic interpretable method called Edges Masking that can be used for any GNN model was designed to explain the underlying mechanisms of GNN model.For five typical epidermal growth factor receptor inhibitors,the interpretable analysis can well clarify their structure-activity relationship,highlighting the rationality of the model.(3)Finally,a free online web server called Kinase Inhibitor Prediction(KIP)was developed to predict the kinome-wide polypharmacology of small molecules(http://cadd.zju.edu.cn/kip).The online webserver has the functions of downloading data set,predicting small molecule’s inhibitory activity against different kinases,and providing the visualization and analysis of prediction results.It aims to help the researchers to quickly screen small-molecular inhibitors against protein kinases in drug design.