Category Archives: Uncategorized

Workshop Announcement: Taming the BEAST in the South Pacific

*** Deadline extended to 7th NOVEMBER for late applications ***

*** Just a few places left. ***

Taming the BEAST in the South Pacific is a comprehensive 5-day workshop to be held on the scenic Waiheke Island, New Zealand from 5 – 10 February 2017.

This workshop will equip researchers with the skills to use BEAST2 software to perform phylogenetics and phylodynamic inferences across a wide range of disciplines through a series of talks by leading experts, lectures and hands-on tutorial sessions. Participants are also encouraged to bring their own datasets for one-on-one discussion and guidance.

Speakers confirmed for the workshop are leading experts in the field:

  • Simon Ho, University of Sydney
  • Alexei Drummond, University of Auckland
  • David Bryant, University of Otago
  • Remco Bouckaert, University of Auckland
  • Tracy Heath, Iowa State University

Taming the BEAST in the South Pacific is hosted by the Centre for Computational Evolution at the University of Auckland, and is modelled after the Taming the BEAST summer school in the Swiss Alps which was organised by the Computational Evolution Group at ETH Zurich. Registration of interest for Taming the BEAST in the South Pacific is open until 25 October 2016. Deadline for late applications is 7th November 2016. Just a few places left, first in first served! See for more information and registration details. Three partial scholarships have been made available for postgraduate students.

Metropolis Coupled MCMC(MC3) works?

19 May 2015 by Remco Bouckaert

Metropolis coupled MCMC (MCMCMC or MC3) allows running an MCMC analysis together with a number of ‘heated’ chains. These heated chains run over a distribution that is adjusted so that it is less peaked than the posterior we want to sample from, which means it is easier for these heated chains to move away from a local optimum. At regular intervals there is the option to switch states between chains (depending on a stochastic critereon), including the chain that samples from the posterior. This is supposed to help explore the sample space more efficient.

To set up an MCMCMC analysis in BEAST, you need to install the BEASTLabs package. The easiest way to set up the XML is by setting it up in BEAUti for a simple MCMC analysis, save the file and edit the XML by

  • replacing the spec attribute in the run element by "beast.inference.MCMCMC".
  • add a chains attribute with the number of chains you want to run.

After this, the XML should look something like this:


When running the analysis, you want to use at least as many chains as there are cores, so that each chain thread can run on its own core. The current implementation is multi-threaded, but does not support multi-processors (yet).

Does MCMCMC work?

The question remains whether it is better to run say 4 individual MCMC analyses and combine results instead of running a single MCMCMC analysis. From what I have seen so far, the BEAST proposals are typically very well tuned to explore tree space, and can handle correlations between various parameters quite well. If a BEAST analysis gets stuck — which shows up by running different chains that seem to converge, but all end up at a different posterior — anecdotal evidence with *BEAST analyses suggest that throwing MCMCMC at it does not solve the problem.

So, there are two criteria on how to judge whether MC3 works or not

  • Can it get us out of local optima, where MCMC by itself has trouble?
  • Can it produce better effective sample size (ESS) per computer cycle?

I can imagine that MC3 works in some cases, and it has been around for ages (notably in MrBayes), but perhaps this is due to the kind of MCMC proposals used, and maybe BEAST analyses do not benefit from MC3. I have not seen an example yet, so if you have a BEAST analysis where MC3 produces better results than MCMC alone, please let me know!