|
In recent years, combination pharmacotherapy has emerged as the standard clinical approach for treating a multitude of complex diseases. Nevertheless, the concurrent use of multiple drugs can lead to interactions that provoke adverse effects in patients, rendering the accurate prediction of potential drug interactions a critical task. Current drug-drug interaction databases are limited in scale, which often results in models learned from these sources struggling to generalization, particularly in cold-start situations where achieving satisfactory predictive performance is challenging. To surmount this obstacle, we propose a drug-drug interaction prediction method based on pretrained models, called PMSE-DDI. We design a pretraining strategy capable of simultaneously learning encoders for both molecules and substructures and use it pretrain on roughly one million unlabeled molecular. Thereafter, the pretrained encoders were transferred to the drug-drug interaction prediction model, and a substructure attention module was developed to pinpoint key substructures within drug molecules. Experimental results in cold-start scenarios demonstrate the superiority of PMSE-DDI in relation to existing methods, as well as the efficacy of the pre-training strategy.
|
