Hi,
I am working with a dataset containing 748 taxa and 24 partitions (each containing sites ranging from 1900 to 173) for a total ~10k sites. BEAST version is 2.1.3 running (with Beagle) on a server containing 48 cores and 64 GB RAM.
As a test, I have had success running BEAST on the 10 MB xml file that includes all 24 site models, a linked clock with relaxed clock log normal model, Yule prior, and a linked tree with a fixed topology (I fixed the topology based on the information on the BEAST wiki).
We have 48 calibrations (including one for the root node) to use. We generated a starting tree using the 'chronos' utility in the 'ape' package in R using a topology we consider reliable (generated with a thorough RAxML run). All 48 interior nodes including the root on the starting tree, fit within the minimum and maximum years (I have checked them visually and programmatically). This was the fixed tree used in the first test run mentioned above.
When I add the 48 calibrations to the xml file, each being a uniform prior, and changed to Calibrated Yule model, BEAST2 loads the file successfully but does not start and hangs indefinitely (well at least for 48 hours). On screen I can see the section where the citations are mentioned. Usually after this point either the starting likelihood is displayed and generations are logged, or, an error is displayed. But I see neither of those.
So to test further, I included only one prior (one for the root node) and I had success: on screen, BEAST2 went past the citation section and calculated the starting likelihood, etc.
Adding the next prior (the next oldest node), BEAST2 gives an error: "Start likelihood: -Infinity after
11 initialisation attempts ..." In the error I can see the starting likelihood of the second prior was -Infinity. Wondering if the second prior was problematic, I used a different prior as the second prior, but got the same error. I have changed the order of the priors, and randomly chose a calibration as the first prior, but apparently, only the prior for the root node seems to work as the first prior.
I have looked at the BEAST FAQ dealing with this error. Based on that, I ran other tests which included changing from uniform priors to say log normal, changing initial substitution and clock rates, changing the branchlengths from MYA to YA instead of changing rates, using different start trees, running BEAST2 with '-beagle_scaling always' option, changing the precision option, doubling the heap space memory for java, even changing the upper and lower bounds on the uniform priors. All lead to the same error: "Start likelihood: -Infinity after 11 initialisation attempts ..." for the second prior.
I would like to hear other people's experience with calibrations in BEAST and if anyone has come across similar issues. Why doesn't beast report anything when all 48 calibrations? What is the issue when I use just two priors and can't initialize? Please let me know if I can elaborate on this problem.
Thank you,
Nitish Narula