While performing co-expression analysis of my RNA seq data, I tried to remove non differentially expressed genes using limma, but I am getting a warning. Please help me out. My RNA seq count data is "hep"

# >hep <- read.csv("hep.csv")
 >log_counts <- log2(data.matrix(hep[,-1]))
> sample <- factor(rep(c("Non.neoplastic","Tumor"), each=21))
> design.mat <- model.matrix(~0+sample)
> colnames(design.mat) <- levels(sample)
> contrast.mat <- makeContrasts(diff = "Non.neoplastic-Tumor", levels = design.mat)
> fit <- lmFit(log_counts, design.mat)
> fit2 <- contrasts.fit(fit, contrast.mat)
> fit3 <- eBayes(fit2)
# Warning:
Zero sample variances detected, have been offset away from zero

The warning seems self-explanatory, is it not? Some of your genes have identical expression values for all samples so the variance is zero. This will make limma's empirical Bayes procedure less effective.

There are three ways you can respond:

1. Filter out unexpressed genes. The warning is quite probably caused by the fact that you have not filtered out unexpressed genes, so you may be including genes for which expression is at the minimum value for all samples.

2. Set robust=TRUE in eBayes, which will make the warning less important.

3. It is just a warning so you can just ignore it.

Later

With the help of ATpoint's comments, it now appears that you are simply inputing RNA-seq counts to limma as if they were expression values without any normalization for library size. This makes your whole analysis incorrect. Probably may of the rows are entirely zero, leading to the warning you see. You instead should follow an appropriate RNA-seq workflow such as voom.