Interactive demo · Statistical Methods for Modern Biology · Chapter 4

Zero-inflated counts: ChIP-Seq bins and the EM algorithm

A zero-inflated model is a two-component mixture: with probability π a bin is a structural zero (no signal can be seen there), otherwise its count comes from a count distribution that can also produce zeros. Which zeros are structural is never observed, so the fit uses EM: the E step gives each observed zero the posterior probability τ = π / P(X = 0) of being structural, the M step updates π and the count parameters. Compare four models on the 256,523 bins of chromosome 22 and see which one reproduces the counts.

Data

Model

Display

Largest count shown

EM (zero-inflated models)

Scoreboard

Try these

EM starts at π = 1/2 and λ = the sample mean (zero-inflated Poisson), or at π = 0.3 and the negative binomial fit to all bins (zero-inflated negative binomial), as in the course's R code. The negative binomial has mean μ and variance μ + μ2/a; for fixed size a its maximum-likelihood mean is the (weighted) sample mean, and a is found by a one-dimensional search. AIC = 2 × (number of parameters) − 2 × log-likelihood.

In the lecture notes