Research On Drug Combination Property Prediction Methods Based On Deep Learning

Compared with single-drug therapy,multi-drug combination therapy has advantages such as better efficacy and less likelihood of developing drug resistance,and has become the standard clinical treatment strategy for complex diseases such as cancer.However,the combined use of multiple drugs may also result in interactions,causing adverse reactions in patients.Therefore,accurate assessment and prediction of drug combination properties,including sensitivity and interactions,are crucial,not only for providing more effective treatment options for patients,but also for reducing the risk of medical accidents.However,it is time-consuming,labor-intensive and cost-expensive to measure the properties of drug combinations through biomedical experiments,so it is difficult to carry out large-scale experiments.In contrast,computational methods,represented by deep learning,are cost-effective and have shown great potential in this field.To address the problem of relatively limited size of drug combination property datasets,which makes it difficult to meet the needs of deep learning models highly dependent on training data,this study provides solutions and strategies from multiple perspectives,and predicts the sensitivity and interaction of drug combinations under both warm and cold start scenarios.The specific work includes:(1)For the drug combination sensitivity prediction task under the warm start scenario,a prediction method based on key substructures of drug molecules is proposed.This method uses substructure decomposition algorithm to extract substructures of drug molecules,and introduces specific encoders to learn substructures.It also constructs a substructure attention module to identify the key substructures that play a more critical role in drug combinations.The feasibility and effectiveness of this method are verified on a real dataset using cancer as an example.(2)For the prediction task of drug-drug interactions under the warm start scenario,a prediction method based on key substructures of drug molecules is proposed.This method assigns embedding vectors to each type of interaction and improves the substructure attention module so that it can pay more attention to the key substructures related to specific interaction types.Experimental results demonstrate the superiority of this method over existing methods.(3)To improve the predictive performance of the drug combination property prediction model built in this study under cold start scenarios,a pre-training strategy that can simultaneously learn molecules and substructures is designed.Based on this strategy,molecule and substructure encoders are pre-trained on unlabeled molecule data and transferred to the drug combination sensitivity and interaction prediction models proposed in this study.Experimental results show that this strategy can effectively improve the predictive performance of the model under cold start scenarios.