# What is the best distribution to model the FPKM values from normalized RNA-Seq data?

I know that the discrete raw counts from the RNA-Seq data are usually modeled by a negative binomial or a Poisson distribution, but what I am working on are the FPKM (Fragments Per Kilobase of transcript per Million mapped reads) values which represent a transformation of the raw values. They are continuous except many actual 0s. What is the probability distribution FPKM values are likely to follow, or are assumed to follow? Even after taking logarithm twice (after replacing 0s with 1, of course), the data does not follow a normal distribution since there are a lot of 0s.

• I think you may want to look into zero-inflated distributions? Commented Aug 31, 2017 at 3:07
• Could you show or provide some sample data to explain with? What have you read about limma, DESeq2, DESeq ?
– llrs
Commented Sep 1, 2017 at 14:45

FPKM/TPM values are generally log-normally distributed. Reference : Gene expression profiling in single cells from the pancreatic islets of Langerhans reveals lognormal distribution of mRNA levels

• Sorry, I added a reference. I'm not sure why mRNA levels follow that distribution but it has been established. As they mention in the paper, it's a fairly common distribution in nature. On the other hand, 3' capture methods of single-cell RNA seq such as DropSeq usually yield a negative binomial distribution, presumably because of the high drop out rates of this method of transcript capture. Commented Sep 14, 2017 at 19:10
• Well done. The answer looks WAY better with the reference. If you would like to push it even further. This type of information you have written in the comment you can add to answer as well to give a complete picture. Commented Sep 17, 2017 at 17:03
• If some data follows a log-normal distribution, does it mean that it would follow normal distribution after you take the log of the values? That's not what I am seeing... Commented Sep 18, 2017 at 22:19
• It should...what does expression look like for a gene you know is expressed in your sample after you log normalize your data? Commented Sep 19, 2017 at 18:20

Cuffdiff is one the tools that deals FPKM data for analysis of differential expression.By default they use the negative binomial with some condition specific parameter to avoid biases. There are some other models also available in Cuffdiff. Check the following documents for more details,