search

beast-dev / beast-mcmc

Bayesian Evolutionary Analysis Sampling Trees
http://beast.community
GNU Lesser General Public License v2.1
192 stars 73 forks source link

extending the epochal discrete CTMC model #1129

Closed jsigao closed 2 years ago

jsigao commented 2 years ago

This pull request contains three commits that extend the epochal discrete CTMC (substitution/dispersal) model. Specifically, this extension allows the transition-probability (P) matrix to be computed correctly along a branch that overlaps with multiple Q-matrix epochs (Bielejec et al., 2014) and multiple average-substitution/dispersal-rate epochs (Membrebe et al., 2019), and the stochastic-mapping function (the endpoint-conditioned uniformization algorithm; Hobolth and Stone, 2009) to work under the epochal model (either epochal matrices or epochal rates or both). A more detailed description about the extension can be found in S1 (S1.1 for the P-matrix computation and S1.2 for the stochastic-mapping function) of Gao et al., 2021.

The first commit (94e91fe) added a getBranchRateModelMapping function to each BranchRateModel class (where the only non-trivial implementation is in the RateEpochBranchRateModel class), imitating the getBranchModelMapping function of the BranchModel classes.

In the second commit (e86792c), this added mapping function is invoked in the SubstitutionModelDelegate class very similarly as the getBranchModelMapping function, allowing the average rate of each Q-matrix epoch that overlaps with a given branch to be computed correctly (instead of assuming that it is the average rate of the entire branch).

The third commit (9e81125) extends the endpoint-conditioned uniformization algorithm to work under the epochal model by (1) dividing a branch into pieces (where each piece only overlaps with one Q-matrix epoch and one average-rate epoch), (2) drawing the state at the end of each piece according to the probability conditioned on the start state of the piece and the end state of the branch, and (3) simulating history over each piece using the time-homogeneous algorithm.

Please let me know if I failed to follow the preferred coding style in this extension or if some of the functionalities introduced here would be preferred to be implemented in an alternative way. Thank you in advance for reviewing this PR.

jsigao commented 2 years ago

Dear @msuchard -- thanks very much for your review! I'll make the requested changes to this PR as soon as possible.

jsigao commented 2 years ago

Dear @msuchard -- sorry for the delay in making the requested changes. As the code-duplication removal affects many files, I've created a new pull request as a revised version of this PR to make the commit history a bit cleaner. To avoid potential confusion, I'll close this PR for now.