Drug-Target Interactions Prediction Via Graph Isomorphic Network and Cyclic Training Method

Predicting drug-target interactions through computational methods holds the potential to provide more reliable candidates for subsequent experimental validation and reduce associated costs. Most methods for DTI prediction have made advancements from two perspectives, improving the accuracy of drug and target representations, and seeking more precise mapping functions between the drug and target spaces. In this study, we propose a model called CT-GINDTI, which prioritizes the optimization of the model training process based on considering aforementioned improvement. CT-GINDTI represents drugs as graphs and utilizes graph isomorphism network to better capture the inherent structural and relational properties of drugs. Additionally, we introduce a cyclic training method to address the imbalance issue between positive and negative samples by selecting more reliable negative samples. To evaluate the performance of CT-GINDTI, we conducted extensive experiments and compared its results with five state-of-the-art methods in the field. The experimental results demonstrate that our proposed CT-GINDTI outperforms these existing methods, showcasing its superior achievement in the prediction of DTIs.