Drug Combination Prediction Based On Network Embedding

The pathogenesis of complex diseases is complex,and if a single drug is used for treatment,it is easy to develop drug resistance.The combination therapy has high efficiency and low toxicity,and is becoming a new treatment for complex diseases.Despite the development of high-throughput drug combination screening platforms,the money and time required to perform drug combination experiments is costly.Therefore,an algorithm is needed to help find potential synergistic drug combinations,further narrowing the scope of drug combination experiments.At present,drug combination prediction models can be divided into four categories: drug combination prediction models based on expression data,drug combination prediction models based on PPI networks,drug combination prediction models based on metabolic pathways,and machine learning drug combination prediction models based on drug similarity.These models have the disadvantages of biased data,poor results,and low available data.This paper proposed a Network Embedding framework in Multiplex Networks for Drug Combinations(NEMNDC),which improved the existing machine learning drug combination prediction model based on drug similarity.Firstly,the importance of each layer of drug similarity network is evaluated by partial drug combination data.Then,the secondorder biased random walk is used for network sampling to obtain the path of random walk,and the training model is generated according to the path.Positive and negative samples;subsequently,the vector representation of the drug node was learned by constructing the Skip-Gram model;finally,the random forest classifier was used for drug combination prediction.In the drug combination forecasting task,this paper evaluates the impact of multilayer network and single-layer network on NEMNDC algorithm.The results show that NEMNDC algorithm can effectively integrate data and has no strong bias for information of each layer network.Among the drug combination results predicted using the NEMNDC algorithm,the first 6 pairs of drug combinations not included in the known drug combination data set have 5 pairs of drug combinations that have been verified to be synergistic by biological and clinical experiments.Analysis of another pair of drug combinations by drug target,GO,and Pathway found that there is a high probability of a synergistic combination of drugs.This paper also applies the NEMNDC algorithm to the drug-drug relationship prediction task.By comparing the INDI,PSL and Mashup algorithms,it is shown that the NEMNDC algorithm is also accurate and applicable in drug-drug relationship prediction.The NEMNDC algorithm not only solves the problem of fixed feature dimension in the previous machine learning drug combination prediction method based on drug similarity,but also constructs the topological information of the multi-layer network in the drug node vector.The multi-layer network importance assessment strategy proposed in this paper can better guide the integration of multi-layer network data.At the same time,using NEMNDC algorithm in drug-drug relationship prediction has obtained good results,indicating its applicability,not only for drug combination prediction,but also for drug-drug relationship prediction and drug relocation.