Drug Repositioning Prediction Based on Multi-source Data and Multi-tasking Models

Abstract

(1) Objectives: In the process of drug repositioning, there are two important problems need to be resolved. One is how to choose the drug candidates and the other is how to select the drug which has fewer side effects than other candidates; (2) Methods: In the paper, a new workflow was created which is combined by integrating the data about drug information to form a dataset and using multitask learning model to process the datasets. In the experiment, we integrated a dataset, in which each drug has 881 features and two labels, and then two machine learning models are used as a screening strategy to predict drug repositioning; (3) Results: In the two methods, MNN(Multitask Neural Networks) and MDBN(Multitask Deep Belief Networks) were both applied for multitasks. After optimization, the MDBN model has the best performance with the mean AUC value of 0.608. It is higher than MNN by 6%. (4) Conclusions: Through the integration of data, the screening range for a certain type of disease can be achieved. The multitask deep model MDBN can be used as a screening method on these datasets. This new workflow can be used as a new idea for drug repositioning to provide more drug candidates. However, the workflow can't fit for some classification which only have several drugs. So the next optimization direction is to mine as much data on these diseases as possible from all perspectives.