The Research On Drug Interaction Prediction Based On Multimodal Representation Learning

Drug-Drug Interaction(DDI)refers to the phenomenon in which one drug changes the pharmacological effect of the other drug when two or more drugs are taken simultaneously or sequentially.Although concomitant use of multiple drugs generally improves treatment efficacy,some combinations can develop unwanted DDIs that can lead to unanticipated Adverse Drug Reactions(ADRs).As a special type of adverse reaction,ADR can affect the health of patients or even lead to death in severe cases.Early detection and identification of DDI can effectively prevent the occurrence of medical malpractice,which deserves the attention of patients,clinicians,and medical researchers.With the continuous increase of data,how to capture the key features of drugs more accurately and effectively from a large amount of complex multimodal data and how to establish a model with strong adaptability and high accuracy is still a problem worthy of research.In this paper,two types of drug interaction prediction are studied,one is Small Molecule Drug(SMD)interaction prediction,and the other is SMD and Biotech Drug(Bio D)interaction prediction.The traditional SMD interaction prediction concatenates different feature vectors without considering the deep interaction and redundant information between features.To address this problem,this paper proposes a DDI prediction method named FM-DDI based on feature reconstruction and multidimensional attention mechanism,which effectively extracts the interaction information between different modalities and removes redundancy for SMD interaction prediction.Furthermore,with the discovery of Bio D,the interaction between two types of drugs,SMD and Bio D,cannot be predicted with traditional DDI prediction methods due to the heterogeneity between SMD and Bio D.To address this problem,this paper proposes a DDI prediction method named Multi-SBI based on multimodal representation learning,which effectively fuses complex information in heterogeneous data for SMD and Bio D interaction prediction.The main research contents and innovations of the paper are as follows:(1)A deep learning method named FM-DDI based on feature reconstruction and multi-dimensional attention mechanism is proposed for small molecule drug(SMD)interaction prediction.The model first inputs and combines multiple drug substructure features to reduce the loss of information caused by a single type of substructure representation.Second,feature reconstruction extracts low-dimensional and informative drug features from heterogeneous data sources(drug molecular fingerprints and association information).Finally,the deep neural network model based on the multi-dimensional attention mechanism assigns high attention weights to the key feature dimensions,thereby effectively capturing key information.Compared to several state-of-the-art drug interaction prediction methods,FM-DDI achieves significant performance improvements.In a case study of psychiatric drugs,7 out of 10 DDIs predicted by FM-DDI with the highest confidence were validated in the latest version of Drug Bank,further demonstrating the effectiveness that FM-DDI extracts and learns drug features to predict SMD interaction.(2)A deep learning method named Multi-SBI based on multimodal representation learning is proposed for interaction prediction of small molecule drug(SMD)and biotech drug(Bio D).Considering the heterogeneous structure and complex network relationship between SMD and Bio D,the model first uses multimodal features to fully represent the structural information and network associations of SMD and Bio D.Second,an under-sampling method based on positive-unlabeled learning,PU-sampling,is introduced to obtain negative samples of high-confidence from unlabeled datasets.Finally,a deep neural network is applied for drug interaction prediction.Experimental results show that multimodal representation learning can represent drug features in heterogeneous drugs more comprehensively,PU-sampling can effectively remove noise from unlabeled samples,and Multi-SBI significantly outperforms other advanced interaction predictions methods.In a retrospective analysis of Drug Bank 5.1.0,14 of the 20 prediction results with the highest confidence were validated in the latest version of Drug Bank,further demonstrating the effectiveness of Multi-SBI to predict SMD and Bio D interactions by learning complex multimodal features of drugs.