TPM, FPKM, WTF?

I (Bob) am dipping my toes back into differential expression modeling for RNA-seq data. Doesn’t everyone love high-dimensional count data? I just can’t get over their standard units: TPM (transcripts per million) is almost OK, but FPKM (fragments per kilo base of transcript per million fragments mapped)? Seriously? We have scientific notation for a reason.

If you’re thinking about bioinformatics…

If you’re going to be looking at RNA-seq data, I would strongly recommend reading the following clear, insightful, and comprehensive Bayesian take on the subject.

Lewin, Bottolo, and Richardson. 2019. Bayesian methods for gene expression analysis. In Balding, Moltke, and Marioni. Handbook of Statistical Genomics, 4th Edition. Wiley.

Yes, that’s Sylvia Richardson, so you know this one’s going to be good. So good that I really wish I’d found it in 2018 when they apparently put the draft out online. But it doesn’t come up in obvious searches for RNA-seq and Bayes on Google and it’s a sad indictment of scholarship that it has zero citations in Google Scholar. This should be among the first things anyone reads as an introduction to microarrays or RNA-seq in bioinformatics! If you have other suggestions in this category, please let me know in the comments.

Since 2018, Shuonan Chen (Columbia systems biology Ph.D. student), Chaolin Zhang (Columbia systems biology professor), and I developed a multilevel Bayesian alternative to rMATS (differential isoform expression with replicates). We’re nearly done with the draft and I’ll announce it here when it’s up on arXiv. We hadn’t read Lewin et al. at the time we developed it, but you can view what we did as a specific instance of the kinds of models they describe in their paper. They even used the trick I learned from Andrew to parameterize the location of two groups as a mean and difference from mean rather than two locations. Even better, the paper goes over how to handle non-uniquely mapped reads, which systems like rMATS can’t handle; my only concern is that they don’t directly address identifiability issues in all these models, which can be challenging.

P.S. My plan’s to write a case study along the same lines as Lewin et al. with more background on the RNA-seq process and simulated data sets with Stan code illustrating the main points, with an extended discussion of Bayesian approaches to non-identifiability. If anyone’s interested in being involved in this, please let me know via email ([email protected]).