We have created a model to integrate several OMICs data, but we realized that the maximum TPM values of RNA-Seq data were so big that had unexpected effects on our results. We hypothesized that this might be due to the negative binomial distribution inherent to this OMIC's layer.
To tackle this circumstance, we proposed to introduce as input the values obtained after using voom() (paper) in our count data. The reason behind that decision was that voom would transform the input data to be more accurately introduced to a linear model. There are, though, several issues/arguments against doing so:
- The results obtained through voom() contain negative values, which do not make sense in our model. I tried to correct that by adding to all the results from a specific sample the absolute minimum value of that sample, but I am not confident whether that is a good way forward.
- Original 0 values (non-expressed) turn into different negative values across samples (-6 in Sample 1, -4 in Sample 2), creating a difference that did not exist to begin with.
- Some sources do not recommend to log-transform count data , including data from negative binomial distributions .
I am not really versed into distribution transformations, so any comment would be appreciated.