Question 18 of 20 Advanced Concept
What information is lost when RNA-seq reads are summarized only at the gene level?
Short answer
Gene counts hide which isoform was expressed, so a switch between isoforms with different functions or lengths can look like no change, or like a false change. They also discard intronic, novel transcript and allele-specific information.
Why
Most genes make several transcript isoforms, and they can differ in function, for example by keeping or skipping an exon that encodes a binding domain. Soneson, Love and Robinson describe three kinds of change. Differential gene expression is a change in total output.
Differential transcript usage is a change in the share of each isoform with total output constant. Differential transcript expression covers either. A gene-level count cannot see usage changes at all.
It can also mislead in the other direction. When isoforms differ in length, a switch from a long to a short isoform lowers the read count with no change in molecule number, and the authors showed this inflates false discoveries in gene-level tests on simple count matrices. Offsets from transcript-level estimates, provided by tximport, remove that bias.
They also found the problem was relatively minor in several real datasets, so gene-level analysis remains a sound default. What it cannot answer is any question about splicing, alternative promoters or alternative polyadenylation. Those need transcript-level or exon-level methods, with the caution that short reads rarely span several junctions, so isoform estimates carry more uncertainty than gene totals.
A strong answer covers
- Distinguish gene expression, transcript usage and transcript expression
- Isoform switching is invisible to gene totals
- Length differences between isoforms can create false gene-level changes
- tximport offsets correct gene-level tests for changing length
- Isoform estimates from short reads are more uncertain
Common mistakes
- Concluding a gene is unaffected because its total count did not change
- Assuming transcript-level estimates are as precise as gene-level counts
- Summing transcript TPMs to get gene counts for a count-based test
The interviewer might ask next
Which analysis would you run to test for differential transcript usage, and what would you report?