A Partial Deconvolution Study Based On Gene Expression Profiles And Prognostic Analysis Of Glioma

With the development of gene sequencing technology,it is easier to obtain massive gene data,and how to find useful information from these data becomes more and more important.In fact,the study of quantification of immune cells from transcriptomics data and the prognosis analysis of glioma are challenging projects which catches the scholars attention.This paper aims to study the collinearity problem of gene microarray data by using a series of simulated data and real data firstly.Our research compares 12 kinds of methods,including ridge regression,principal component regression,partial least-squares regression,LASSO,adaptive LASSO,elastic net,the generalized elastic net,CIBERSORT,EPIC,quadratic programming,DSA and Decon RNASeq method.We can summ up the ridge regression method not only can effectively solve the multicollinearity problem caused by error,also has a higher accuracy by comparing with the real proportion of cells and error variance.On the other hand,this article uses public glioma patient data sets from CGGA and TCGA to set up two Cox proportional hazard models.The one is based on clinical and cell proportion data and the other is based on clinical and gene expression data.This paper use LASSO and ISIS to select variable.We studied the PH assumption of the model based on the filtered variables and add interaction term with time to the variables that not consistent with PH assumption.We can conclude that the model based on clinical and cell proportion data is better by comparing C-index,Time-dependent ROC and model robustness.