Given a matrix of limma voom normalized genes, I want to do Survival Analysis of a particular gene X.

The design matrix applied on voom was achieved regarding recurrence or no recurrence. Yet I have three disease groups A, B and C. When I filtered patients only with disease C and did DGE over this patients.

This gene X appeared overexpressed in patients that showed cancer recurrence in group C.

Now I need to convert the matrix into a matrix underexpressed / overexpressed / normal so that I can do survival analysis to show indeed that for patients with higher expression of this gene the cancer returns.

How should I convert this matrix into this nominal shape?

Possibilities:

1. I understand that limma voom results in negative, neutral (near to 0) and positive, so negative means underexpressed, near to zero normal and positive mean overexpressed for all genes?

2. Inside the same gene X do histogram, analyze bi modality and set threshold negative/ neutral /positive by regression?

3. Get the log2 foldchange and log(adj.p values) of each gene and define threshold in the same way vulcano plot is designed?

1. No, limma-voom computes the logCPM for each gene and then estimates observation-level weights. The logCPM value is proportional to the underlying gene expression but has nothing to do with under/normal/over-expression.
2. What if it's not bimodal? Why does it need to be trichotomized?
3. The logFC is a measure of differences between groups, not the measure of each individual's gene expression.

If you want to use gene expression as a predictor for a survival analysis, the naive thing to do is to take the logCPM value of that gene and use it as a predictor for the survival analysis. If you already have the gene you care about, then that's all you need to do. If you don't already know what gene to use, then you could use the CRAN glmnet package to identify genes that appear to be related to survival.

But all of this is off-topic for this site, and if you have any further questions you should go to biostars.org, where it is on-topic.