Research On Prediction Of Drug-Target Interaction Based On Feature Combination And Transfer Learning

The prediction of drug-target interactions is a critical step in drug discovery and drug repositioning research technology,which can provide scientific basis at molecular level for drug discovery and drug repositioning.And it can not only effectively save human and financial resources invested in biological experiments in the process of research and development,but also shorten the period of drug development.It has important scientific research significance and practical engineering application value.Gene expression profile can reflect the level of gene expression in human cells.Different drug stimulations and gene knockouts can cause abnormal expression of different genes.Based on the data of gene expression profiles of drugs and targets,the thesis combines with the low-order feature combination characteristics of FM model and the ability of DNN model to learn high-order feature representation from the point of feature combination of drugs and targets.DeepFM model is used to predict drug-target interaction.On this basis,a RNN-CNN hybrid model is proposed.RNN can learn the combinatorial relationship between drug and target features,and CNN can learn the linear combination of drug and target features at the same point.Finally,for some models with insufficient predictive performance,transfer learning method of weighted domain adversarial neural networks is used to train the model.The knowledge learned from other cell lines is transferred to the target cell line according to different degrees to avoid negative transfer and to improve the robustness of the prediction model.In the experiment,gene expression profiles and drug-target interaction pairs in LINCS and DrugBank datasets are used to generate positive samples of known drug-target interaction pairs and negative samples of drug and non-target pairs.Then the positive samples and negative samples are combined as data sets for model training.The experimental result shows that the proposed feature combination model can effectively obtain better accuracy of prediction by studying the feature combination relationship between drugs and targets than the general machine learning method.In addition,this thesis proposes using the idea of transfer learning to improve the prediction performance of some models.The transfer learning method can effectively improve the prediction robustness of the model to a certain extent and provide a research direction for the drug-target interaction of small sample set.