Introduction to Bioinformatics and Computational Genomics

Week 10Transcriptomics and differential expression

From a count table to a list of genes that changed, and the three sources of noise that stop you believing the first answer.

Questions this week answers

  • When do you normalise for gene length, and when is doing so a mistake?
  • The treatment did nothing at all. How many of your 10,000 genes still come out significant?
  • The gene with the biggest fold change is at the top of your table. Is it your result?

By the end of this week you can

  • Name the three sources of variation in RNA-seq counts
  • Say when to normalise for library size and when for gene length, and when not to
  • Explain why 10,000 tests at p < 0.05 need correcting
  • Read a log2 fold change and an MA plot
0 of 5 done

  1. Why a p value of 0.05 is not a promise

    Set the treatment to have no effect at all, then raise the gene count. Every bar that appears is a gene you would have written up.

    Nothing changed, and you would still publish 100 genes. That is not a subtle statistical caveat; it is most of a results section.

    Genes you would report1000 real, 100 noise
    False positives100from 2,000 genes that did not change
    Share of your list that is real0.0%
    Threshold per test0.0500
    A bacterial genome is a few thousand; a human transcriptome is tens of thousands.
    The p value below which you would call a gene significant.
    Assumed to be detected every time, which is generous.
    Correction
    Things to try0 of 3

    The 10,000-gene example is from the week 10 DGE statistics lecture.

What should survive this week

  • Three sources of variation: technical, shot noise, and biological. Biological dominates, and it is why counts follow a negative binomial rather than a Poisson.
  • Normalise for library size always. Normalise for gene length only when comparing different genes; for differential expression the length cancels.
  • 0.05 x 10,000 = 500 significant genes from a treatment with no effect. This is why an adjusted p column exists.
  • log2 makes fold change symmetric around zero. Sort by adjusted p, not by fold change: low-expression genes produce enormous ratios from noise.