| Background:The rapid development of high-throughput sequencing technology and the dramatic reduction in testing costs have provided an unprecedented opportunity to screen and identify omics biomarkers related to the prognosis of complex diseases at the population level.Statistically,the identification of biomarkers related to disease prognosis can be included in the category of survival analysis.The prognosis of complex diseases is the result of the joint action of multiple genes,but it is not a simple accumulation between them.Multiple genes often interact in a biological network,which controls the prognosis and outcome of the disease.However,at present,almost all statistical methods of identifying genes related to the prognosis of diseases remain the level of a single gene,and existing polygenic methods often ignore the complex network relationships between multiple genes,which is difficult to explain the network mechanisms of disease prognosis.Therefore,a network regression model for survival analysis should be constructed,and then achieve the transformation from "identifying a single gene marker for complex disease prognosis" to "identifying a gene network marker for complex disease prognosis".Gene interaction networks are usually composed of nodes representing genes and edges representing the functions or interactions between different nodes.Each gene node and edge depict specific biological meanings that cannot be ignored.Since the effect of the overall network includes "node effects" and "edge effects",both the changes of nodes and edges in the network can affect the prognosis of diseases,To explore the association between networks and the prognosis of complex diseases,it is necessary to obtain both gene expression data and survival data in the same sample.However,obtaining gene expression data through whole gene expression profiling sequencing technology is often expensive.In recent years,transcriptome-wide association studies(TWASs)integrating genome-wide association studies(GWASs)and expression quantitative trait loci(eQTLs)provide a theoretical framework to obtain large-scale gene expression data through data integration.Methods:In this study,we integrated the disease prognosis network regression model with TWAS.Firstly,for a specific gene in the network,using genotype and gene expression data in a small sample eQTL study to construct a gene expression prediction model to estimate the effect of genotype on gene expression.Furthermore,this estimated3 was substituted into GWAS to obtain the predicted gene expression(network node).At the same time,point mutual information(PMI)was used to characterize network edges,and network nodes and network edges were included to construct a Cox proportional hazards model for network regression in TWAS(CoNet)simultaneously to detect the association between specific gene networks and survival outcomes.CoNet was compared with the traditional TIGAR method and CPNT(cox productive hazards model based on product moment for network regression in TWAS)comprehensively.Statistical simulations were used to evaluate its scientific ity and effectiveness,and the real breast cancer data from UK Biobank were analyzed and evaluated for practicality.It should be noted that the CPNT model used product moment to characterize linear correlation as network edges.Its comparison with the CoNet model will fully demonstrate the ability of the CoNet model to capture relationships between complex and multi-type network nodes.1.In statistical simulation experiments,detailed statistical simulations were designed based on GEUVADIS and UK Biobank.It includes two types of simulation scenarios:(1)prespecified nodes and edges in the network;(2)Randomly selected nodes and edges in the network.Four network modes were designed for each simulation scenario,including only nodes in the network had effects,only edges in the network had effects,both nodes and edges in the network had effects and the effecting node was on the effecting edge,and both nodes and edges in the network had effects but the effecting node was not on the effecting edge.Under each network mode,consider different network node correlation modes,including the combination of sine and quadratic,quadratic,sine and linear.At the same time,comprehensively investigate the statistical performance and robustness of the CoNet model under different sample sizes(5000,10000 and 20000),different censoring rates(0.1,0.3 and 0.5),and different gene expression prediction models.2.In the real data analysis,using the genotype and follow-up data of breast cancer patients in UK Biobank,7 biological networks potentially related to breast cancer in KEGG were selected,and the genes contained in each network were matched with the eQTL to obtain the final network.Further,we obtain the predicted expression value(network node)of each gene in the specific network based on the TWAS framework.The network edges were calculated in PMI and PM separately to complete the real data analysis.Considering that nodes and edges were often highly correlated in TWAS network regression,and the commonly used multiple test method Bonferroni correction is too strict,this study used false discovery rate(FDR)to adjust the P-value and set the FDR significance threshold to 0.05.Results:The simulation results showed that both CoNet and CPNT were able to control type Ⅰ errors and remained robust around the significance level of 0.05 under various simulation scenarios such as correlation patterns between different network nodes,different censoring rates,and different sample sizes.When testing for node effects,CoNet and CPNT had comparable power.When testing for edge effects,the power of CoNet and CPNT increased with the increase of sample size and the decrease of censoring rate.When the correlation mode between nodes was nonlinear,CoNet had higher power than CPNT.At the same time,the difference in testing efficiency between the two methods is related to the patterns of nonlinear relationships.In the case of the combination of sine and quadratic(xk=sin2xl)and quadratic relationships(xk=0.1xl2),CoNet’s testing performance is significantly higher than CPNT,while in the case of sin relationships(xk=sinxl),the difference in testing performance decreases.Applying the two models to breast cancer data,CoNet identified four genes with eight edges,CPNT identified four genes without edges,and the two methods identified three same genes(CDK6,DTX3L,and PTGS1).Conclusions:Under different sample sizes,between-node correlation patterns,and censoring rates,CoNet can effectively identify effecting nodes and edges,effectively control type I errors,and has higher power compared to TIGAR and CPNT.Even under high censoring rates and missing data,CoNet remains robust.In addition,CoNet can be used for large-scale dataset analysis,such as UK Biobank. |
