Research On The Prediction Method Of Drug-side Effect Frequency Based On Frequency Value Propagation

Determining the frequency of drug side effects is a key issue in drug development and drug risk-benefit assessment.Accurately and efficiently predicting the frequency of side effects of drugs is of great significance for reducing the probability of drug safety problems and protecting the health of patients.The frequency of side effects of drugs is often determined in randomized controlled clinical trials,but this method has limited performance and high cost.With the increase of drug clinical trial data,it is possible to study drug side effects based on data-driven computational methods.Existing prediction methods of drug-side effects mainly studies the association between drug and side effects.In contrast,this thesis aims to study the problem of predicting the frequency of drug-side effects.According to the frequency of side effects in drug development clinical trials,conduct quantitative analysis and comparison for each side effect of each drug.Our thesis studies and proposes a collaborative propagation model RPDSF（Rating Propagation model for predicting Frequencies of Drug–Side effects）based on high-order similarity network of drug-side effect.The model builds biological entity similarity networks based on known drug-side effect frequency information,and deduce the frequency of potential drug side effects through the process of high-order collaborative propagation of known frequency information in the network.Considering that different scoring functions may lead to different similarity scores and model performance when building similarity networks,this thesis constructs similarity networks in various ways,and further proposes an improved method of similarity networks based on neighborhood learning to improve the accuracy of model prediction.In this thesis,the effectiveness of the method is verified on real drug-side effect frequency data,and the performance of the collaborative propagation model based on different similarity networks is analyzed,among which,the RP₀（69）4)9)2))model based on neighborhood learning achieves the minimum prediction error with RMSE=0.5798and MAE=0.4150.At the same time,this thesis also compares with other related work based on the same data set.Compared with the method adopted by Galeano et al.who firstly conducted this study,the RP₀（69）4)9)2))model and RP₀（69）4)9)2))model proposed in this thesis achieved 54.90%and 56.12%decrease in RMSE,and decrease 54.49%and57.19%in MAE.In addition,compared to the current best performing related work,the RP₀（69）4)9)2))model in this thesis also decreased by 10.75%and 14.84%in RMSE and MAE,respectively.The experimental results show that the prediction performance of the proposed model is better than other models.