Bayesian inference of time trees

Introduction

• Bayesian inference tries to obtain a probability distribution rather than a single best fit c.f. maximum likelihood
• This is typically obtained using numerical approaches such as Markov Chain Monte Carlo
• For example, Metropolis sampling involves the following
  • Make a change to the parameters
  • If the new parameters result in a better fit, change to them
  • If the new parameters result in a worse fit, move to them with a probability based on the drop in fit
  • Repeat many times

Strict and molecular clocks

• In phylogenetic inference, Bayesian approaches have become popular to fit molecular clocks to sequence data
• Two commonly used platforms
  • MrBayes
  • BEAST
• Both use the same underlying principles, but the models are parameterised a little differently
  • BEAST has a wide range of 'demographic models' that can be used as prior distributions for the branch lengths
• The input is also different
  • BEAST has an XML input format
  • MrBayes uses the Nexus format