Prediction of synergistic effects of drug combinations for cancer treatment based on multilevel information fusion

Drug combination therapy is essential in cancer treatment, significantly enhancing the survival rates of cancer patients. In previous studies, researchers have investigated various methods to predict drug synergy, such as utilizing drug chemical information and protein-protein interaction network (PPI) to obtain features of both drugs and cancer cell lines. However, there has been comparatively little research on the information regarding drug molecular structures and the associations between cell target proteins and cancer cell line genes. Therefore, this paper sets out to investigate this aspect. We propose a model that utilizes a three-layer Graph Isomorphism Network (GIN) to capture the structural information inherent in the SMILES representations of drug molecules. Additionally, we employ a graph attention networks (GAT) to learn features of PPI network. Finally, we integrate drug features, cancer cell line target protein features, and gene features, and utilize XGBoost model to predict the synergistic effects between drugs. We evaluate our model on the DrugComDB and Oncology-Screen datasets, and experimental results demonstrate its superior performance. On the DrugComDB dataset, our model achieves F1, Recall, and ACC scores of 0.738, 0.718, and 0.768, respectively. Similarly, on the Oncology- Screen dataset, our model demonstrates F1, Recall, and ACC scores of 0.810, 0.825, and 0.790, respectively. Our model outperforms several established models across all metrics. We believe it can effectively identify combinations of anticancer drugs.