Multi-granularity Self-attention Model-based Prediction Method Of Drug Activity And Drug Target Association

Natural products have been the main source of new drugs,but more than two-thirds of them have no identified target.Identification of drug activity and the relationship between drug and target is a key problem in drug discovery and drug relocation.The traditional biological experiment method has the problem of high cost and time consuming.The computation-based method has the advantages of high speed,low cost and strong interpretation,providing a powerful tool for drug research and development.In this thesis,drug activity prediction and drug-target association prediction are studied.The main research work includes: firstly,a representation learning method based on multi-granularity self-attention model is proposed.Guided by the "drug analogy text" and based on the sequence data of drugs and targets,the initial feature vector representation of drug and target sequence data is obtained through pre-training,and multiple complex data are no longer required as input.The self-attention model was used to learn the drug and target feature representation from the drug and target sequence of various particle sizes,and the drug property prediction and drug-target association prediction problems were solved based on the obtained feature representation.Secondly,two methods for drug property prediction and drug-target association prediction were proposed based on the representation learning method of multi-granularity self-attention model.Compared with the existing methods,the proposed model has better results in seven real data sets,including BACE,anti-osteoporosis and ESOL.At the same time,through further experimental verification with real data,the model proposed in this thesis can successfully predict the activity of 17 unknown drugs from the TOP20 unknown drugs.In addition,based on the research method in this thesis,an online platform named Inflam Nat(www.inflamnat.com)was designed and implemented for the study of antiinflammatory natural products,which contains a database of physicochemical properties,cellular anti-inflammatory biological activities and molecular targets of currently discovered natural products.Two prediction tools based on multi-granularity selfattention designed for natural products(a)predict anti-inflammatory activity of natural products and(b)predict unknown drug-target associations collected from the database are provided.The platform realizes the retrieval function of natural products,which can retrieve the basic information,physical and chemical properties,anti-inflammatory biological activity,and corresponding target information of the included natural products,and can predict the anti-inflammatory activity of drugs online,as well as the correlation between drugs and targets.