EBT: a statistic test identifying moderate size of significant features with balanced power and precision for genome-wide rate comparisons

Abstract

Supplementary data are available at Bioinformatics online.

Improvement in prediction of prostate cancer prognosis with somatic mutational signatures

“…For example, the prognosis of PCa could be further improved with new stratification criteria. The small number of PCa cases and the general low somatic gene mutation rates in PCa could also have led to the dramatic low power 25 . Therefore, no solid conclusions could be drawn before more objective-targeted studies are performed with enlarged size of cases and observation of elongated period of survival.…”

“…To compare the somatic gene mutation frequency between prognostic groups, a matrix was prepared to record the mutations of all the genes for each case, followed by counting the number of cases with mutations for each gene in each group. Both Chi-square test with Benjamini & Hochberg correction and EBT were used for rate comparisons, and a False Discovery Rate (FDR) or p value < 0.05 was set as the significance level for Chi-square or EBT test respectively 24 , 25 .…”

“…The genes were filtered when any of the following criteria was met: (1) the mutation rates in both groups were lower than 5%; (2) the absolute difference between the mutation rates of two groups was lower than 5%; (3) the significance of Chi-square test without FDR correction was higher than 0.1. Both TopN and mRMR strategy were also used for feature selection and model comparison 25 , 26 . For TopN strategy, the top N genes with smallest p values (EBT) for mutation rate comparison were selected as the features 25 .…”

“…Both TopN and mRMR strategy were also used for feature selection and model comparison 25 , 26 . For TopN strategy, the top N genes with smallest p values (EBT) for mutation rate comparison were selected as the features 25 . The mRMR software package was downloaded, installed and used for mRMR feature selection 26 .…”

“…An R package, 'e1071', was used for training Support Vector Machine (SVM) models using each training dataset ( http://cran.r-project.org ). For each training-testing experiment, the training dataset was used for both kernel selection and parameter optimization as described previously 25 . Four kernels, including 'radial' (Radial Base Function, RBF), 'linear', 'polynomial' and 'sigmoid', were individually tested for the best-optimized parameters with a 10-fold cross-validation grid search strategy.…”