Muli-Task Dual-view Model for Drug Combination Analysis: In Vivo and In Vitro Perspectives

Drug-drug interactions (DDIs) and drug compatibility interactions (DCIs) are both critical components of drug combination (DC), essential for ensuring safe and effective therapeutic strategies. DDIs typically occur in vivo, while DCIs occur in vitro. However, most existing drug combination prediction models focus solely on either DDIs or DCIs, lacking a comprehensive approach to predict DC in vivo and in vitro perspectives, which can lead to incomplete understanding and management of DC risks. To address this problem, we propose a multi-task dual-view model for drug combination (MDDC) that simultaneously predicts DDIs and DCIs by using pre-trained molecular sequence and graph representations. MDDC utilized the pretrained models ChemBERTa-2 and MolCLR to capture diverse view molecular features and fine-tune them in a multi-task learning framework. This approach demonstrated notable performance improvements, with AUROCs of 0.914 for DDI task, and 0.826 for DCI tasks on independent testing dataset. MDDC offers a more comprehensive understanding of potential drug interactions in medication guidance, thereby reducing adverse effect risks and enhancing therapeutic safety.