Multi-Scale Feature Fusion For Drug-Drug Interaction Prediction

Drug-Drug interactions refer to adverse side effects that occurred when two or more drugs were taken.The adverse side effects of drugs are likely to lead to other diseases or even death.Therefore,it is crucial to accurately predict drug-drug interaction.The fundamental methods for predicting drug-drug interaction are the traditional laboratory-based methods.These lines of methods are labor-intensive,time-consuming,and costly,and can cause combinatorial explosion problem when all drugs were taken into accounts.With the development of high-throughput sequencing and deep learning technologies,a large amount of biomedical data has been extracted,which provides valuable information for drug-drug interaction task.The major challenge is that how to extract the potential valuable information from it.In this paper,we studied the potential adverse reactions between drugs with knowledge graph and graph neural network technologies based on the biomedical data from various sources.The main research of this paper are as follows:(1)Considering that most studies only focus on drug-drug similarity network or the integration of multiple similarity networks,but rarely extract the correlation between nodes in different networks and high-order semantic information,we proposed a Graph DDI model,a knowledge-aware graph attention network model,to predict drug-drug interaction.Firstly,we collected the data from Drug Bank and KEGG,the widely used databases,and then converted it into the knowledge graph.Secondly,we used the knowledge representation model to learn the representation of nodes and relations in the knowledge graph,and then used them as the initial vectors of nodes and relations of the subsequent network.After that,we designed the graph attention propagation module to learn the importance of different neighbors for target node.After multi-layer iterations,the nodes can integrate local topology structure and high order connected semantic information.Moreover,we evaluated the Graph DDI model on binary-class drug-drug interaction task,and the experimental results demonstrate that Graph DDI achieved the better performance.(2)Considering that previous works either focused on the structure information of the drugs without considering the rich semantic information related to drugs,or utilized knowledge graph(KG)with rich bio-medical information without considering drug molecular structure information,and they do not consider the synergistic effect between the drug chemical structure and KG,thus limiting its predictive capability.Moreover,most state-of-art works consider is the presence of an interaction between drugs by considering the DDI prediction as a binary classification task while ignoring the significant research on specific types of adverse reactions between drugs.Considering the above limitations,we propose a novel multi-scale feature fusion(MUFFIN)model,a deep learning framework for DDI prediction using drug chemical structure and bio-medical KG.We designed a bi-level cross strategy that can jointly learn the fusion representation of internal(chemical structure)and external(KG)features of the drugs from convolutional neural network(CNN)-based cross-and scalar-level perspectives.The bi-level architecture can effectively combine the multi-modal features through the multi-granularity feature fusion process,thus improving the capability of DDI prediction.Moreover,we evaluated the MUFFIN model on three different DDI prediction tasks,namely,binary-class,multi-class,and multi-label tasks.Experimental results demonstrate that MUFFIN achieved the best performance on three tasks,thus supporting the significance of the combination of chemical structure and knowledge features from KG.