gene-level versus transcript-level analysis

Question

Traditionally, RNA-seq data was quantified on gene level. Newer methods quantify on transcript/isoform level. For example, Kallisto only outputs transcript-level abundances. From the DESeq2 vignette:

A newer and recommended pipeline is to use fast transcript abundance quantifiers upstream of DESeq2, and then to create gene-level count matrices for use with DESeq2 by importing the quantification data using the tximport package.

Is it just for consistency and/or simplicity that the values are converted to gene level? Would it better to proceed with transcript-level table or does that violate some DESeq2 (or similar tools) assumptions?

Answer 1

There are a variety of reasons people use gene-level quantitations.

1. Transcript-level differences are difficult to biologically interpret. Let's be honest, few groups are likely to put in the work required to determine what these might mean. Most genes have at least some level of characterization, so it's much more biologically tractable to think in terms of gene-level changes. Granted, "genes" are largely artificial constructs and in an ideal world one would always think in terms of transcripts, since they're what's really changing. But we don't live in an ideal world.

2. Statistical power increases with counts, so if you're splitting your counts among a handful of isoforms you could be missing something. In practice this isn't a big problem, since there are typically obvious major isoforms, but that won't always be the case.

3. "Mo transcripts mo problems", to channel the bioinformatics notorious B.I.G. It's common practice for people to use a subset of transcripts when using tools like Kallisto and Salmon, since the results usually end up being better. But which subset of transcripts should you use? Those in refseq? Gencode basic? A subset of one of those? Do we really understand all the consequences of choosing one over the others? (No, we don't) We all have our own preferences, but this isn't a solved problem.

Answer 2

To add to the list that Devon Ryan outlined (or perhaps to elaborate on point 2?):

Although Salmon/Kallisto/RSEM are the more accurate in their transcript quantification than the methods they superseded, transcript level quantification is still not as accurate as gene level quantification (see the tximport paper, which would also be the tool i'd recommend for getting gene level estimates from the transcript-level output of these tools).

That said, the tximport authors do claim that considering the transcript-level quantification in your gene-level analysis improves the accuracy of the gene-level analysis.