| The task of predicting the interactions between drugs and targets plays a key role in the process of drug discovery.In the past decades,there have been many biological studies dedicated to identifying drug-target interactions.However,biological experiment-based research methods are time-consuming and expensive.Therefore,developing effective computational methods to identify drug-target interactions is important.Most of the existing graph neural network-based methods are based on isomorphic graphs.These models only obtain information from drug-related networks or target-related networks with directly connected nodes.They cannot be applied to heterogeneous graph bioinformatic networks with complex relationships.Therefore,they not only fail to capture the higher-order dependencies in heterogeneous graphs,but also ignore the potential semantic information in heterogeneous graphs.Meanwhile,most of these models are trained using only labeled data.The labeled data is expensive and limited.This not only underutilizes the information implied in the unlabeled data but also limits the generalization ability of the models.In this paper,We propose a heterogeneous graph representation-based learning drug-target interactions prediction model(HGDTI)to address the above problem.The model uses metapaths to model the semantic structure in heterogeneous graphs.We use graph attention networks and semantic-level attention mechanisms to obtain representations of nodes.To improve the model’s generalization ability,we construct a self-supervised graph comparison learning task using mutual information.After experiments,our model is shown to exhibit the best performance on the standard public dataset of DTIs.In heterogeneous networks,different meta-paths are potentially related.They reflect different perspectives of the same object.If each meta-path is treated as an isolated semantic data resource and the correlations among them are disregarded,sub-optimality in both the metapath-based embedding and final embedding will result.To address this problem this paper proposes a collaborative knowledge distillation-based drug-target interactions prediction model(CKDDTI).In this paper,we use the simplified molecular input line entry system sequence of the drug and the amino acid sequence of the target as the nodal features of the drug and the target.More specifically,we model the knowledge in each meta-path with two different granularities: regional knowledge and global knowledge.Then the mutual information is used as a distillation metric.We learn the meta-path-based embeddings by collaboratively distilling the knowledge from intra-meta-path and inter-meta-path simultaneously.Finally,we obtain the ultimate representation of the drug and target for association prediction.In this paper,experiments are conducted using a drug-target standard dataset.The results demonstrate that the model in this paper can accurately perform the interaction prediction. |
