Research On Drug-Drug Interactions Algorithm Based On Multimodal Graph Neural Network

The synergistic effects of multiple drug combinations have been widely used to treat diseases,especially complex diseases such as cancer.However,combination medications can also cause side effects,such as antagonism and even toxicity.Therefore,accurate identification of drug interactions is critical to public health and patient life safety.At the same time,as the size of drug databases grows,detecting large numbers of drug interactions using traditional methods,such as biological or clinical trials,requires a lot of manpower,money and long trial cycles.In order to solve the above problems,the current researchers tend to adopt machine learning methods.In general,such methods use the topology generated by known drug interactions to predict unknown interactions.In order to improve the predictive performance,different drug characteristics are also used as feature representation to describe the drug.It’s still a challenge to effectively integrate the multi-source information of drugs and improve the predictive ability.In this paper,the multimodal graph neural network method is used to predict drug interactions.The main contents are as follows:(1)The drug-drug interactions can be transformed into a link prediction problem,with known drug interactions as edges in the drug interaction graph.The main purpose of the study was to predict the presence of edges between pairs of drugs with unknown interactions based on known edges.(2)Existing prediction algorithms based on graph neural network mainly learn node feature representation on a single fixed graph network topology,but this structure cannot reflect the complete graph structure information,which will limit the expression ability of the model.To solve this problem,drug interaction graphs under different modes were constructed,and then the low-dimensional representations of each drug under different modes were obtained,which using the graph neural network message passing mechanism.Finally,the multiple representations of drugs in different modes were integrated into a unified representation to predict the drug interaction.(3)A multimodal graph neural network(MMGNN)algorithm was proposed,and the experiment results on Drug Bank data set showed that the performance of this algorithm for drug interaction prediction was better than that of the current algorithms.In order to further explore the predictive power of MMGNN,this paper makes a case study from two aspects.The experiment results demonstrated that MMGNN can effectively predict unknown drug interactions.