Adversarial Enhancement Models Of Multilayer Graph Convolutional Networks For Virtual Drug Screening And Online Platforms

Since traditional drug development and discovery is supposed to be a time-consuming and expensive task with a high failure rate.Deep learning based methods have been deployed to model virtual screening in recent years.And since drug molecules can be naturally represented as graph structures,where atoms and bonds between atoms are nodes and edges of the graph.Hence,it is natural to apply the deep graph learning for training the model of the ligand-based virtual screening.However,the deep graph learning always shows its low performance due to poor generalization when applied in drug activity regression prediction with a small scale training set.At the same time,The drug developers have to spend much time and effort in code development and the construction of deep learning based model,rather than in the drug performance analysis and drug mechanism.To address these issues,we propose a drug virtual screening algorithm based on multilayer graph convolutional network(VSMGCN).First,the molecular features extracted by multilayer graph convolutional network are projected into two subspaces to improve the local smoothness of the deep learning model.Second,the perturbations of the adversarial direction are assigned to one of the feature subspaces,which becomes less vulnerable to a small bias and improves the generalization performance of the deep graph model.We applied our method on 12 GPCR datasets,and confirmed that our method achieves superior performance in terms of prediction accuracy and top-k screening of biology activity in comparison to five commonly-used deep graph based methods.Finally,we design an online drug virtual screening platform to perform our proposed method,where the following functionalities can be implemented through the interaction between the frontend and back-end:(1)select or submit datasets online and divide them into training sets and test sets to accommodate the variety of requirements.(2)select the feature extraction module and the feature processing module online to generate a new algorithm model.(3)The online parameter tuning and deep graph model training for the drug screening can be achieved by the visualization of the loss function and metric.Multiple experiments have been conducted on this online drug virtual screening platform and verified that the deep graph model for drug virtual screening can be effectively constructed and debugged online on this platform,which can significantly improve the efficiency of drug virtual screening in a simpler and easier way.