Prediction Of Drug-related Phenotypes Based On The Constructed Phenotype-gene-drug Heterogeneous Network

Understanding the relationship between disease phenotypes and drugs is an impor-tant issue in network medicine.With the continuous enrichment of phenotypic and drug data,scientists have proposed various approaches to explore biological information from these data.Among the existing methods,network-based ones show their unique advan-tages.However,most of the considered networks were centered on disease networks,and rarely considered networks that were composed of three or more biological entities.There-fore,it is inappropriate to directly use these methods to identify the relationship between disease phenotypes and drugs.Here we present a new method called correlation prediction algorithm based on the shortest path between disease phenotype and drugs in heterogeneous networks(RSP)to identify the relationships between disease phenotypes and drugs.The method inte-grates disease phenotype data,gene data and drug data,which are extracted from three different databases,based on the obtained information,we can obtain data about the interactions between disease phenotypes and gene,drug and protein targets.After that,a heterogeneous network containing interactions between three different biological entities is constructed.The degree distribution,connectivity and functional similarity are ana-lyzed to explore the differences between the sub-networks of the heterogeneous network and the random networks.Next,we used the Dijkstra algorithm to calculate the short-est path length between disease phenotypes and drugs from the heterogeneous network,which was used to improve the correlation coefficient between the phenotype class and the drug class.Subsequently,the correlation coefficient scores between the drugs and the phenotypes were predicted.Given the significant difference in the correlation coef-ficient scores,we used the Youden’s index to determine the appropriate threshold,and delete the interaction between the drug and the phenotype with correlation coefficient less than the threshold to achieve a more accurate prediction effect.Finally,the AUC value,the Youden’s index,and the number of successfully predicted drug-phenotype pairs were used to evaluate the ability of the method to identify the relationships between the phenotypes and the drugs.We find that the RSP method outperforms some other meth-ods(CIPHER、ProphNet、the Shortest Path(SP)).In addition,we also show that our method is robust to gene networks,and the result indicate that our method is useful in the prediction of phenotypic targets of experimental drugs.