Skip to content

Latest commit

 

History

History
121 lines (75 loc) · 3.22 KB

5-DeSeq2Intro.md

File metadata and controls

121 lines (75 loc) · 3.22 KB
Raw

R Notebook

Introduction to DeSeq2

How does deSeq work?

We want to compare between two biologically distinct conditions and try to figure out how the gene expressions differ between these two conditions.

For instance, we have two untreated and treated drug conditions.

We have biological replicates and multiple technical replicates from these samples.

We get a count matrix from these experiments.

Counts are always in

  • Dynamic range

  • doesn’t follow Normal distribution

  • in whole numbers

Then what distributions best represent these data?

  1. Poisson distribution?

Why poisson distribution

  • you fit this distribution when the number of cases is large but the probability of event happening is low.

  • in RNA-seq selecting mRNA from a large number of RNA molecules sort of represent this situation.

But then why does poisson distribution not work?

The mean and the variance of the poisson distribution has to be the same!

However, genes tend to have higher variance than mean.

For low mean counts, high variance but high mean counts, low variance.

AS A RESULT DeSeq2 uses a NEGATIVE BINOMIAL distribution

you can account for the extra variability with this distribution.

What does DeSeq2 do?

For each gene…

  • Estimate size factor and dispersion
  • Fit linear model
  • Hypothesis test (p-value)

Estimate Size Factor

There are two bias we have to take account to: Library Size and Library Composition

  • Library size : sequencing depth might be a problem

  • Library composition

only gene D is differentially expressed

Gene untreated treated
A 2 10
B 4 12
C 6 20
D 30 0
  1. Calculate Geometric mean (geometric mean is more robust to outliars)
Gene untreated treated pseudo ref untreat / ref treat / ref
A 2 10 sqrt(20) = 4.47 0.45 2.24
B 4 12 6.93 0.58 1.73
C 6 20 10.95 0.55 1.83
D 30 0 0 0 0

You get the median of untreat/ref and treat/ref

median of untreat / ref = 0.5, median of treat/ref = 1.78 = normalization factor

divide by size factor

Gene untreated_norm treated+norm
A 4.016 5.65
B 8.032 6.74
C 12.05 11.24
D 60.24 0

Now these are normalized for library composition

Estimate the dispersion of each gene

measure the spread and account for variability between replicates

assumes gene with similar expression has similar dispersion

Step 1: gene dispersion estimate Maximum Likelihood Estimator

Step 2: fit a curve to gene-wise dispersion estimates

Step 3 : Shrink gene-wise dispersion estimates towards values predicted by the curve

Step 4 : if the variation is too high, deSeq2 will ignore and say there are unaccounted variation

Fit a linear model

deSeq2 fit a Generalized Linear Model