Transforming RNA-seq data- (tpms) - what is the best practice for dealing with zeros for the purpose of deconvolution?

Question

[nancy]

Using Cibersort to deconvolve bulk RNA-seq data to specific cell-types based on expression values from bulk RNA-seq. I have a series of samples, and their individual TPMs as computed by Stringtie. Each column is a sample, each row is a gene. Since cibersort only accepts non-zero values, I am trying to understand the best way to transform those tpms which = 0.

Are any of the below more preferred/accurate?

1. Treat all samples and all rows the same- Add a small (0.0001) value to all the rows for all samples
2. Treat all samples the same for those rows which = 0 - Add a small (0.0001) value to only those rows with 0
3. Treat each sample differently - Find the smallest non-zero value for each sample, and add that value toall the rows for that sample.

Thanks

[Kristoffer Vitting-Seerup]

What do the authors of Cibersort do in their paper?