Research On Human Oral Bioavailability Prediction Based On Graph Neural Networks

Human oral bioavailability is one of the most important pharmacokinetic properties in human oral drug development.Accurately predicting the human oral bioavailability of drug candidates can reduce the resource consumption in the development of new oral drugs.Currently,machine learning algorithms are often combined with specific computational methods or based on expert-defined molecular descriptors to build predictive models.This approach not only require manual extraction of molecular descriptors,which increases the workload,but also do not bring new insights and ideas to oral drug development.In recent years,graph neural networks have been widely used in molecular property prediction because of their ability to naturally model molecular structure information.However,previously proposed graph neural networks do not adequately consider the interactions between atoms and bonds,which limits the ability of the model to represent molecules.In addition,the perception range of molecular structure is different for graph neural networks with different depths.The fusion of molecular hidden representations of different depths by sample adaptation may improve the molecular representation capability of graph neural networks.Therefore,this thesis proposes improved methods to develop more accurate and efficient models for human oral bioavailability prediction from two major aspects: atomic and chemical bond information interaction and model structure definition,with the following main work and contributions:(1)A Bond Message Absorption Network(BMANet)based on directed graphs is proposed.A Bond Message Absorption Mechanism is proposed to increase the interaction between atoms and bonds to obtain a better representation of molecules.Using Scaling Self-Attention,the model focuses on important features,scaling up the values of important features while reducing the values of non-important features to improve molecular representation ability.By comparing with four machine learning models and eight graph neural network models,it is demonstrated that BMANet outperforms other methods and possesses strong interpretability.By testing on a public graph neural network molecular property prediction dataset,it is demonstrated that BMANet is equally applicable to other molecular property prediction tasks and outperforms other models.(2)A sample-based adaptive Dynamic Depth Graph Neural Network(DD-GNN)is proposed.It enables graph neural networks to adaptively fuse molecular hidden representations of different depths based on molecular structure information,and while filtering noisy information to improve molecular representation ability.Test results on publicly available datasets demonstrate that the performance of all four graph neural networks combined with DD-GNN improves to different degrees,proving the versatility of DD-GNN for graph neural networks.(3)Combining BMANet with DD-GNN to develop a more robust human oral bioavailability prediction model.The results show that DD-BMANet has a lower error in quantitative predicting human oral bioavailability.In most cases,DD-BMANet also outperformed BMANet in qualitatively predicting human oral bioavailability.This thesis explores the possibility of using graph neural networks to predict human oral bioavailability.A Bond Message Absorption Network and A sample-based adaptive Dynamic Depth Graph Neural Network are proposed.A large number of experiments demonstrate the advanced,interpretability of the proposed method in the prediction of human oral bioavailability.And it is also applicable to other molecular property prediction tasks,providing a more efficient and accurate prediction model for new drug development.