thebench.bio #1 Ecosystem for the Bio-Tech Industry
All RNA-seq questions

Question 11 of 20 Intermediate Interpretation

Why can transcript-length differences complicate comparisons across genes within the same sample?

Short answer

A longer transcript produces more fragments at the same number of molecules, so raw counts rank long genes too high. Length must be corrected to compare gene A with gene B in one sample, but largely cancels when comparing one gene across samples, unless isoform usage changes.

Why

Library preparation fragments each RNA molecule, and each fragment can become a read. A 10 kb transcript and a 1 kb transcript present at the same number of copies give roughly a tenfold difference in reads. Raw counts therefore mix abundance with length.

Measures such as TPM divide by transcript length so that genes inside one sample can be ranked fairly. The best-practice review by Conesa and colleagues makes the distinction directly: length correction is not needed to compare the same gene across samples, but it is needed to rank expression levels within a sample. Length also has a hidden effect on statistics.

Long genes collect more reads, so they have more power to be called differentially expressed. The goseq authors showed the proportion of genes called significant rises with transcript length, which biases later enrichment tests toward categories full of long genes. The comfortable assumption that length cancels across samples breaks when a gene switches between a long and a short isoform between conditions.

Its effective length changes, and tximport's average transcript length offsets exist to handle exactly that case.

A strong answer covers

  • Reads scale with abundance times length
  • Within-sample ranking needs length correction, across-sample comparison of one gene usually does not
  • Long genes have more power to be called significant
  • Isoform switching changes effective length between conditions

Common mistakes

  • Claiming one gene is more highly expressed than another from raw counts
  • Dividing by gene length before running a count-based differential expression test
  • Ignoring length bias when interpreting enrichment of a differential expression list