Research On Multi-attribute Drug Classification Based On Neural Network

In recent years,with the development of high-throughput screening and the accuruulation of drug informatics data,drug research and development(R&D)has entered a new phase.Drug repositioning,as one of drug R&D strategies,can be used to find new indications,new mechanisms,or new side effects through big data mining algorithms,and significantly reduce the cost,risk and time of drug R&D.At mean time,multi-source drug infonnatics data(chemical property,clinical property,and pharmacological property,etc.)bring huge opportunities for multi-dimensional analysis of drug mechanism and drug repositioning.This dissertation proposes a computational analysis framework based on drug informatics data for the two core problems of predicting new indications of FDA approved drugs and new drug-target interaction prediction.Firstly,a large number of drug perturbation and gene knockout transcriptome data are obtained from LINCS database,and a complete sample space is constructed through pre-processing steps.Then a two-layer perceptron model with Softmax is designed and extended to the deep neural network for systematically exploring new drag-target interactions and new indications of FDA approved drugs.Resulting from the multi-view/multi-modal drug attributes,this research establishes a fusion system for homogeneous/heterogeneous data.On the one hand,the drug-target data crossing multiple cell lines are integrated by designing a variable depth data cube.On the other hand,a domain-specific feature extractor based on adversarial strategy is defined to get a latent space among multi-source drug informatics data.The above works improve the accuracy and confidence of new drug indications and new drug-target identifications,and the scalability of the model has been enhanced.The main research includes the following four aspects:(1)The two-layer perceptron model for drug treatment attribute classification.The prediction of new therapeutic attributes is modeled as a multi-label classification task,since a drug usually has multiple therapeutic attributes.Based on the transcriptomics data of 480 FDA approved drugs in PC3 cell line,a two-layer perceptron model with Softmax as non-linear activation is built for learning high-level representations and predicting new therapeutic attributes.And the efficiency of prediction has gotten obvious improvement.(2)The deep neural network for drug-target interaction classification.Transcriptome data of 480 drug perturbations and 4,363 target knockouts in PC3 cell line and the existing drug-target interactions in DrugBank database are used as gold standard.A deep neural network with dual data channels is constructed.In order to get better classification performance and more precise decision boundary,the distribution of sample space and the penalty weight of the objective function are rebuilt.(3)The variable channel-convolutional neural network for drug-target interaction classification.In order to effectively fuse data from different views(i.e.transcriptome data of drugs and targets in different cell lines),this research aligns drug-target pairs in different cell lines and designs a variable depth data cube.Then,the variable channel-convolutional neural network is constructed to model the correlation between different view data for achieving fusion of seven cell line data and promoting the performance of drug-target interaction classification.(4)Combining adversarial strategy and multi-task learning for drug treatment attribute classification.In order to effectively integrate multiple data and label sources,this research uses adversarial strategy to map heterogeneous data with several feature patterns to the same feature space with domain-specific information rather than domain-invariant information.The bidirectional long short term memory network is adopted for fusing multiple domain-specific representations,and the multi-task learning framework is established for joint learning among classes.Last,new indications for FDA approved drugs and massive small molecule candidates are predicted with high confidenceThe above four models range from single source to multiple sources.Not only the accuracy of drug repositioning gets improved,also some effective solutions for multi－source data fusion have been proposed.What’s more,the pervasiveness and efficiency of model provided significant technical support for the development of drug research.