Prediction Of Drug Target Interactions Based On Heterogeneous Information Network Representation

Drug development is a high-cost,long-time-consuming,wide-ranging,and lowefficiency development project.The drug development process is complex due to the vast array of drug types and intricate functional mechanisms.Moreover,there are various types of drug networks,such as drug-target interaction networks,drug-path association networks,etc.Therefore,finding undiscovered drug associations from drug networks has essential research value.The vast majority of drug-target interaction predictions currently model the drug relationship as a homogeneous information network.However,drug networks naturally constitute a heterogeneous information network due to the diversity of their types and associations,and link prediction is a hot problem in heterogeneous information network analysis.The issue of link prediction refers to whether there is an edge between nodes in a heterogeneous information network.Existing deep learning methods have shown exemplary performance in link prediction.However,these methods ignore the mutual disturbance of different edge types in the network.Specifically,models trained on other type labeled edges may provide conflicting conclusions for unlabeled edges.In addition,it is difficult to collect enough labeled edges for one type of edge in the drug network and use that type of edge to train the model.Therefore,weakening this type of disturbance is very challenging for link prediction.In order to solve the above problems,this paper proposes a prediction framework based on heterogeneous information network representation,called Type Resilient Tag Model(TRTM).It consists of three main components: a structural characterizer,a typeexpert constructor,and a resilient predictor.First,the structural characterizer is responsible for learning edge representations for link prediction in drug-targets,and extracts corresponding structure features from different edge types by enhancing the type structure features.The type-expert constructor cooperates with the structural characterizer to remove the type-common features and retain the structure-specific features of specific types to generate type-expert models for different edge types.Ultimately,the resilient predictor develops a unified resilience predictor mechanism to fuse different conclusions from unknown edges of different edge types to get the best prediction results.The TRTM model proposed in this paper not only learns different types of knowledge in the drug heterogeneous information network in the training phase to generate corresponding types of experts,but also develops a resilient predictor mechanism to aggregate the prediction labels of multiple types of experts in the prediction phase.Therefore,TRTM has strong resistance to edge-type disturbances during training and prediction stages.Furthermore,TRTM provides a general framework to explain edge-type disturbances during link prediction in drug heterogeneous information networks.The structural feature in the model can also be easily replaced with other heterogeneous information network representation models,which greatly enhances the applicability of the TRTM model.