Prior Information-dependent Differential Network Analysis Method And Its Application

Objective:Understanding the changes of gene regulation networks in different state is of great significance to elucidate the mechanism of disease and to identify the driver genes involved in the disease.In this study,we evaluated the performance of pDNA by simulation study and apply it to construct the differential network between the Luminal-A and Basal-like breast cancer.We aim to identify hub nodes in the network and provide evidence for clinical diagnosis,treatment and development of targeted drugs.Methods:We introduced the basic principles and steps of the prior information-dependent differential network analysis method and compared it with PNJGL and DiffCorr.In the simulation,we sample size n=50,100,200;number of perturbed nodes m=4;proportion of differential edges?₁=0.3,0.5,0.7;proportion of view-specific differential edges?₂=0.1to generate six views of data with p=100 genes.By calculating precision and recall,the Precision-Recall rate curve was drawn.The performance of the three methods with each combination of n and₁?was evaluated using precision-recall curve.In real data analysis,we apply pDNA using gene expression and methylation data to construct the differential networks between Luminal-A and Basal-like breast cancer and identify hubs.The selected hub nodes were included in the logistic regression model,and the 10-fold cross validation was carried out to draw the ROC curve.Results:On the whole,we observed that pDNA substantially outperforms PNJGL and DiffCorr.The performance of pDNA in constructing differential network is affected by network density and the sample size.With the increase of sample size and network density,the performance of pDNA becomes better.For the Luminal A and Barsal like breast cancer,we construct a weighted network including 562 nodes and 1,041 edges.We consider the top 10 genes with the largest degree as hubs.Seven of them?YWHAG,PRKCG,CHUK,ATF2,PIK3RT,TSC1,NR4A1?have been reported to be associated with breast cancer.AUC was 0.85?0.81,0.90?.Conclusion:According to the simulation study,we observed the competitive performance of pDNA.The usage of the non-paranormal graphical model can relax the normality assumption.In addition,pDNA can incorporation prior information into differential network inference and integrate information across multiple datasets.Hub nodes in the estimated differential networks between the Luminal-A and Basal-like breast cancer rediscover cancer-related regulator genes and contain some prediction information.FGF13 may be associated with classification and prognosis of breast cancer,and its role requires further study.