KGECDR: drug response prediction based on element-aware knowledge graph attention network
Ge Zhang, Qiqi Si, Xinxin Li, Lijun Zhang, Hui Yang · 2024
Rapid screening and accurate application of anticancer drugs is an important issue in the era of precision medicine, which is crucial for achieving personalized treatment. Many computational methods are currently being developed to predict drug responses. However, most methods ignore the richness of information about other biological entities (proteins, tissues, etc.) associated with drug-cell lines and lack interpretability. To address this problem, we propose a drug response prediction method based on an element-aware knowledge graph attention network. This method can model drugs and cell lines at a fine-grained level, thus identifying potential drug responses. First, KGECDR models the combination of relationships in the knowledge graph related to drugs and cell lines as elements related to drug cell line response (CDR). Second, KG-level embeddings of drugs and cell lines are learned through relation-aware aggregation. Third, element-level embeddings of drugs and cell lines are obtained for KG-level embeddings using element-aware message aggregation based on known CDR. Finally, the obtained embedding of drugs and cell lines can be used to predict the drug response to the cell line. Experiments demonstrated our method can be effective for anticancer drug response prediction.