| With the rapid development of the Internet,the computer toxicology has become an important means of drug toxicity prediction.It has gradually emerged as important solution to vitro predicted potential compound toxicity due to its rapidness,accuracy,non-pollution,and low cost.The traditional machine learning model cannot effectively identify the complex relationship between molecular structure and toxicity.With the growth of Deep Learning,it has been a hot research topic to construct molecular toxic prediction model based on the graph neural networks.In this paper,a molecular toxicity prediction algorithm based on the graph homogeneous and attention network is proposed,where the chemical molecular string is first mapped into the molecular structure graph.Then molecular embedding features are extracted via the graphic homogeneous network.Furthermore,the neighbor atoms features are utilized to enhance the feature representation via the graph attention netowrks.Therefore,the structural properties can be learned to improve the accuracy of toxic prediction.On the baise,a multi-task-based molecular toxicity prediction algorithm via the graph homogeneous and attention network is proposed,where a unified deep learning model is constructed by using multiple relevant tasks.Furthermore,the feature extraction part of this model based on homogeneous and attention network can be trained in a unified way,meanwhile the output layer of each task is trained separately.Hence,multi-task mechanism can enchance the generalization ability of model to further yield improvement in the accuracy of toxic prediction.Finally,multiple experimnets on twelve TOX21 data sets were conducted to compare with the signle task method and other multi-task methods,and comparison shows our proposed method outperforms the competitors,verifying the advantage and feasibility of our methods. |
