A Study On Knowledge-Augmented Deep Learning Approaches For Drug Repositioning

Drug repositioning is a promising approach to identify new therapeutic uses for existing drugs,which can substantially reduce the time and cost of drug development and enhance the efficiency of the process.Existing computational drug repurposing methods can be broadly classified into four categories: classical machine learning models,matrix-based methods,network-based methods,and deep learning methods.These approaches have limitations in handling data sparsity,missing data,multiple association types and interpretability,restricting their application and dissemination.To overcome these limitations,we proposed a novel method named UKEDL-DR,which employs knowledge graph embedding,pre-trained models and recommendation system for drug repositioning.UKEDL-DR uses the PairRE method for knowledge graph embedding,which utilizes paired vector representations of relationships to handle complex relationships between entities such as drugs,diseases,and proteins,which may exhibit many-to-many or one-to-many associations.The CReSS method is employed to extract drug features,which uses deep contrast learning technology to extract the characteristics of drug molecules,and can capture the influence of molecular chemical environment.DisBERT for disease feature extraction is a pre-trained language model based on Transformer that captures complex relationships among biomedical entities within diverse and extensive biomedical texts.By introducing knowledge graphs and pre-training,the problem of data sparsity has been resolved.Finally,a recommendation system is introduced to enhance the interaction between drugs and disease features and improve the accuracy of drug repositioning.We evaluated UKEDL-DR on benchmark datasets,and the results demonstrate its superiority other models in terms of AUC,AUPR,precision,F1,and recall rates at various top-k levels.For instance,UKEDL-DR achieves a recall rate of 0.9584 at the top-500 level,significantly higher than those of the LR,RF,and DeepDR models,highlighting its exceptional performance in identifying potential therapeutic uses for existing drugs.Finally,the model interpretability is increased through the visualization of knowledge graphs.The UKEDL-DR method presented in this study combines knowledge graph embedding,pre-trained models and recommendation systems,and offers a comprehensive and practical approach for drug repositioning.