A Study On Prediction Methods Of Drug Properties Based On Network Diffusion Algorithm

Drugs are a kind of chemicals that have curative effects and improvements for certain diseases of humans or other organisms.Familiarity with the characteristics of drugs is beneficial to the usage of drugs in clinical medicine,can also improve the efficacy and reduce the cost of drug development.Therefore,more and more researchers begin to study new information of drugs based on the validated of drugs,so as to obtain more drug-related information.This article focuses on the two characteristics of drugs,including drug side effect and drug ATC code.The network diffusion algorithm and machine learning algorithm are adopted to predict the potential side effects of drugs/chemicals for all classes in the first level of ATC system.When studying drug side effects,the drug and side effects are paired as samples,thereby modeling the problem of predicting drug side effect as a binary classification problem.Based on the chemical/drug interaction network,the restart random walk algorithm is applied to extract negative samples with different quality,which adopts the known drugs in one side effect as seed nodes.Then five features on fingerprint,structure,ATC code,text information and target protein are extracted to represent all positive and negative samples.These features are learned by three machine learning algorithms: random forest,support vector machine and neural network.Obtained models are evaluated by ten-fold cross-validation.Finally,compared with the models with random selection strategy and FIRE strategy,the strategy proposed in this paper is obviously superior to the other two methods.In the drug/chemical characteristics,the ATC code is very representative.This article conducts a prediction study on the drugs in the first level of the ATC classification system.First,the validated drugs in one class are used as seed nodes,the Laplacian thermal diffusion algorithm is executed on a chemical/drug interaction network.Each candidate drug/chemical is assigned a heat value.To improve the reliability of candidate drugs/chemicals,three tests are performed on candidate drugs/chemicals: including permutation test,association test and functional test.Drugs/chemicals that passed these three tests are considered as new members of the class in the first level of the ATC system.