Tree Search with Profile Parsimony

Getting started

A companion vignette gives details on installing the package and getting up and running.

Once installed, load the inapplicable package into R using

library('TreeSearch')

In order to reproduce the random elements of this document, set a random seed:

# Set a random seed so that random functions in this document are reproducible
suppressWarnings(RNGversion("3.5.0")) # Until we can require R3.6.0
set.seed(888)

Scoring a tree, and conducting a tree search

Here’s an example of using the package to conduct tree search with profile parsimony. You can load your own dataset, but for this example, we’ll use a simulated dataset that comes bundled with the TreeSearch package.

data(congreveLamsdellMatrices)
my.data <- congreveLamsdellMatrices[[10]]
my.phyDat <- phangorn::phyDat(my.data, type = 'USER', levels = c(1, 2))

We then need to prepare our dataset for Profile Parsimony by calculating the information loss implied by each additional step in each character. Ideally we’d pick a high value of precision (> 1e+06): this takes a while, but only needs doing once for each dataset of a given size.

my.prepdata <- suppressWarnings(PrepareDataProfile(my.phyDat, precision = 4e+04))

To start analysis, we need to load a starting tree. We can do this at random:

tree <- TreeTools::RandomTree(my.phyDat, root = TRUE)

Or using a neighbour joining method, to start at a reasonably good tree:

tree <- TreeTools::NJTree(my.phyDat)
par(mar = rep(0.25, 4), cex = 0.75) # make plot easier to read
plot(tree)

Let’s calculate this random tree’s parsimony score, then search for a better tree:

ProfileScore(tree, my.prepdata)

## [1] -279.5769

better.tree <- ProfileTreeSearch(tree, my.prepdata, EdgeSwapper = RootedTBRSwap)

##   - Performing tree search.  Initial score: -279.576861720899

##   - Final score -279.576861720899 found 0 times after 100 rearrangements.

The parsimony ratchet (Nixon, 1999) is better at finding globally optimal trees. ProfileRatchet is a convenient wrapper for the underlying function Ratchet.

# Longwinded approach:
better.tree <- Ratchet(better.tree, my.prepdata, searchHits = 10, 
                       searchIter = 100, ratchIter = 5,
                       swappers = list(RootedTBRSwap, RootedSPRSwap,
                                       RootedNNISwap),
                       InitializeData = ProfileInitMorphy,
                       CleanUpData = ProfileDestroyMorphy,
                       TreeScorer = ProfileScoreMorphy,
                       Bootstrapper = ProfileBootstrap)

# Equivalent, but less typing!
RootedSwappers <- list(RootedTBRSwap, RootedSPRSwap, RootedNNISwap)
better.tree <- ProfileRatchet(better.tree, my.prepdata,
                              swappers=RootedSwappers,
                              searchHits=10, searchIter=100, ratchIter=5)

Let’s see the resultant tree, and its score

attr(better.tree, 'score')

## [1] -309.1615

par(mar = rep(0.25, 4), cex = 0.75) # make plot easier to read
plot(better.tree)

The default parameters may not be enough to find the optimal tree; type ?Ratchet to view all search parameters.

View the results

In parsimony search, it is good practice to consider trees that are slightly suboptimal (Smith, 2019).

Here, we’ll take a consensus that includes all trees that are suboptimal by up to 1.5 bits. To sample this region of tree space well, the trick is to use large values of ratchHits and ratchIter, and small values of searchHits and searchiter, so that many runs don’t quite hit the optimal tree. In a serious study, you would want to sample many more than the 25 Ratchet hits (ratchHits) we’ll settle for here, probably using many more Ratchet iterations.

suboptimals <- ProfileRatchet(better.tree, my.prepdata, 
                              swappers = list(RootedTBRSwap),
                              returnAll = TRUE, suboptimal = 5, 
                              ratchHits = 25, ratchIter = 500, 
                              bootstrapHits = 15, bootstrapIter = 450,
                              searchHits = 10, searchIter = 100)

The consensus of these slightly suboptimal trees provides a less resolved, but typically more reliable, summary of the signal with the phylogenetic dataset (Smith, 2019):

par(mar=rep(0.25, 4), cex=0.75)
plot(my.consensus <- ape::consensus(suboptimals))