This paper addresses the challenge of predicting drug-target interactions (DTIs) by modeling them on signed heterogeneous networks. The authors propose a novel framework called SHGNN-DTI, which integrates information from drug-drug interactions (DDIs) and protein-protein interactions (PPIs) to enhance the prediction accuracy. The framework is designed to handle both signed bipartite networks and two-level signed heterogeneous networks, where interactions between drugs and targets are categorized into positive or negative based on their pharmacological effects. The SHGNN-DTI model employs a three-module architecture that includes a signed graph neural network (SHGNN) for node embedding, a discriminator for predicting interaction signs, and modules for incorporating DDIs and PPIs. The model is trained using different training modes, including cooperative and independent modes, and can share parameters across modules. Experimental results on two datasets from DrugBank show that the SHGNN-DTI framework outperforms existing methods, achieving high accuracy and F1 scores. A case study on two breast cancer drugs, Goserelin and Epirubicin, demonstrates the model's ability to discover new DTIs, with seven out of the top-10 predicted interactions supported by literature. The study highlights the effectiveness of the SHGNN-DTI framework in predicting DTIs with signed information, providing valuable insights for drug discovery and development.This paper addresses the challenge of predicting drug-target interactions (DTIs) by modeling them on signed heterogeneous networks. The authors propose a novel framework called SHGNN-DTI, which integrates information from drug-drug interactions (DDIs) and protein-protein interactions (PPIs) to enhance the prediction accuracy. The framework is designed to handle both signed bipartite networks and two-level signed heterogeneous networks, where interactions between drugs and targets are categorized into positive or negative based on their pharmacological effects. The SHGNN-DTI model employs a three-module architecture that includes a signed graph neural network (SHGNN) for node embedding, a discriminator for predicting interaction signs, and modules for incorporating DDIs and PPIs. The model is trained using different training modes, including cooperative and independent modes, and can share parameters across modules. Experimental results on two datasets from DrugBank show that the SHGNN-DTI framework outperforms existing methods, achieving high accuracy and F1 scores. A case study on two breast cancer drugs, Goserelin and Epirubicin, demonstrates the model's ability to discover new DTIs, with seven out of the top-10 predicted interactions supported by literature. The study highlights the effectiveness of the SHGNN-DTI framework in predicting DTIs with signed information, providing valuable insights for drug discovery and development.