PDTDAHN: Predicting Drug-Target-Disease Associations using a Heterogeneous Network

Background: Disease is a major threat to life, and extensive efforts have been made over the past centuries to develop effective treatments. Identifying drug-disease and disease-target associations is crucial for therapeutic advancements, whereas drug-target associations facilitate the design of more effective treatment strategies. However, traditional experimental approaches for identifying these associations are costly and time-consuming. Numerous computational models have been developed to predict drug-target, drug-disease, and disease-target associations. However, these models are designed individually and cannot directly predict drug-target-disease associations, which involve interconnections among drugs, targets, and diseases. Such triple associations provide deeper insights into disease mechanisms and therapeutic interventions by capturing high-order associations. Objective: This study proposes a computational model named PDTDAHN to predict drug-targetdisease triple associations. Method: Six association types retrieved from public databases are used to construct a heterogeneous network comprising drugs, targets, and diseases. The network embedding algorithm Mashup is applied to extract features for drugs, targets, and diseases, which are then combined to represent each drug-target-disease association. The classification model is trained using LightGBM. Results: Cross-validation on eight datasets demonstrates the high performance of PDTDAHN, with AUROC and AUPR exceeding 0.9. This model outperforms previous models based on pairwise association predictions. Conclusion: The proposed model effectively predicts drug-target-disease triple associations.