So, this is a classification problem, where the outcome Y is binary: NOF and COF are associated with one or more gene expressions (TRUE) or not (FALSE). Conventionally, your null hypothesis is that there is no association. Your task is to develop the appropriate test statistic to estimate \hat{Y} and make a judgment, before hand is what you will expect as a value of that statistic sufficient to make the call on rejecting the null hypothesis and what statistical confidence interval would be useful.
Unless there's a quantitative variable hiding in your data somewhere ANOVA , there's no variance or standard deviation to analyze. You have observations of patients and a series of T/F measures, NOF, CAF, TUMOR, GENE\_i, ... GENE\_n that the patient either expresses or fails to express. If you load
library(edgeR)
edgeRUsersGuide()
you'll see that the documentation talks about ANOVA-like functions only.
BTW: library names in R are case-sensitive.
A workflow for gene expression will get you started with populating patient gene expression with DESeq2. Once you have those
library(edgeR)
edgeRUsersGuide()
and carefully review § 4.1 RNA-Seq of oral carcinomas vs matched normal tissue to see how generalized linear models glm are tailed in edgeR for your domain problem.