Heterogeneous Graph Neural Network For Predicting Biomolecular Interaction Algorithms

Interaction between biological molecules is a complex biological mechanism that involves nonlinear dynamics among many molecules in biology.How to better integrate the complex relationships between these different molecules is a challenging problem.LncRNAs and miRNAs are two types of very important non-codingRNAs,and lncRNAs are non-codingRNAs longer than 200 nucleotides.On the other hand,the functions of more than 99% of lncRNAs are currently unknown,which is a huge contradiction between the important biomedical value of lncRNAs.How to quickly infer the function of lncRNA is one of the hot researches in the field of bioinformatics.The interactions between lncRNA and protein are the most important molecular mechanism of lncRNA,and studying the interactions between lncRNA and protein is a key way to explore the function and molecular mechanism of lncRNA.Traditional experimental methods for predicting lncRNA-protein interactions are time-consuming and expensive.Many machine learning-based computational methods have been developed to predict lncRNA-protein interactions.However,discovering the interactions between lncRNAs and proteins remains a challenge.miRNAs are noncoding single-stranded RNA molecules with a length of about 22 nucleotides,which are involved in many biological regulatory processes and are closely related to a variety of diseases.They have become candidate molecular targets for new disease diagnosis and treatment.The known associations between predicted miRNAs and diseases are still very limited,so it is very urgent and challenging to build computational models to predict the relationship between potential miRNAs and diseases.Our research focuses on the prediction of lncRNA-protein interactions and the prediction of miRNA-disease associations.And we construct two graph neural network models for predicting lncRNA-protein interactions and predicting miRNA-disease associations.To better integrate heterogeneous network information,in the first part of this study,we propose a multi-order heterogeneous graph neural network model HGNN?LPI for LPI prediction.First,a heterogeneous network of lncRNA-protein interactions is constructed based on multi-order neighborhoods,including lncRNAdisease associations,protein-disease associations,lncRNA-lncRNA interactions,protein-protein interactions,and lncRNA-protein interactions;then,apply a heterogeneous graph neural network to capture deeper semantic and topological properties;after extracting features from the heterogeneous network,the comprehensive experiment for different classifiers is discussed to select the optimal classifier model.We compare the proposed model with the most used homogenous network representation model and heterogeneous network representation respectively,the effectiveness of our model is demonstrated.Finally,a web server(http://hgnnlpi.com/)is developed to provide to maximize its availability.In the second part of this study,we propose a model HGCN?MDA for miRNA disease associations prediction based on hierarchical heterogeneous graph convolutional networks.The main work includes: firstly,the associations between miRNAs and diseases,miRNAs and miRNAs,and diseases and diseases are collected;then we calculate the attributes of nodes in the graph and use them as indicators to stratify the network,and further construct a multi-layer network to extract the global embedding and local embedding of nodes,and we integrate different embeddings according to their importance.Finally,we use a fully connected layer to build a prediction model.The results show that the model has good performance.