Drug-Target Interaction Prediction via Deep Multimodal Graph and Structural Learning.

Drug-target interaction (DTI) prediction speeds up drug repurposing, accelerates drug screening, and reduces drug design timeframe. Previous DTI prediction frameworks lack consideration for the multimodal nature of DTI, advanced feature representation techniques, and generalizability on unseen drugs and proteins. Therefore, we propose a novel framework that combines a multimodal graph neural network with direct, molecular-level structural learning via model ensembling. We use a multimodal biomedical that contains drugs, proteins, diseases, and pathways, all of which have meaningful feature embeddings generated via language models or knowledge graphs. We also employ a structural learning module that exploits molecular-level information and runs independently from the graph. Lastly, the graph and structural modules are combined, forming the optimal prediction. Our proposed framework outperformed multiple benchmark DTI frameworks on real-world datasets. After testing on an independent dataset, we conclude our framework is generalizable to unseen drugs and proteins. Our model can be easily extended to other biomedical link prediction problems, such as drug-drug interaction.