MaximizeParsimony: Find most parsimonious trees

Description

Search for most parsimonious trees using the parsimony ratchet and TBR rearrangements, treating inapplicable data as such using the algorithm of Brazeau2019;textualTreeSearch.

Tree search will be conducted from a specified or automatically-generated starting tree in order to find a tree with an optimal parsimony score, under implied or equal weights, treating inapplicable characters as such in order to avoid the artefacts of the standard Fitch algorithm @see @Maddison1993; @Brazeau2019TreeSearch. Tree length is calculated using the MorphyLib C library Brazeau2017TreeSearch.

Usage

MaximizeParsimony(
  dataset,
  tree,
  ratchIter = 7L,
  tbrIter = 2L,
  startIter = 2L,
  finalIter = 1L,
  maxHits = NTip(dataset) * 1.8,
  maxTime = 60,
  quickHits = 1/3,
  concavity = Inf,
  ratchEW = TRUE,
  tolerance = sqrt(.Machine[["double.eps"]]),
  constraint,
  verbosity = 3L
)
Resample(
  dataset,
  tree,
  method = "jack",
  proportion = 2/3,
  ratchIter = 1L,
  tbrIter = 8L,
  finalIter = 3L,
  maxHits = 12L,
  concavity = Inf,
  tolerance = sqrt(.Machine[["double.eps"]]),
  constraint,
  verbosity = 2L,
  ...
)
EasyTrees()
EasyTreesy()

Value

MaximizeParsimony() returns a list of trees with class multiPhylo. This lists all trees found during each search step that are within tolerance of the optimal score, listed in the sequence that they were first visited, and named according to the step in which they were first found; it may contain more than maxHits elements. Note that the default search parameters may need to be increased in order for these trees to be the globally optimal trees; examine the messages printed during tree search to evaluate whether the optimal score has stabilized.

The return value has the attribute firstHit, a named integer vector listing the number of optimal trees visited for the first time in each stage of the tree search. Stages are named:

seed: starting trees;
start: Initial TBR search;
ratchN: Ratchet iteration N;
final: Final TBR search. The first tree hit for the first time in ratchet iteration three is named ratch3_1.

Resample() returns a multiPhylo object containing a list of trees obtained by tree search using a resampled version of dataset.

Arguments

dataset: A phylogenetic data matrix of phangorn class phyDat, whose names correspond to the labels of any accompanying tree.
tree: (optional) A bifurcating tree of class phylo, containing only the tips listed in dataset, from which the search should begin. If unspecified, an addition tree will be generated from dataset, respecting any supplied constraint. Edge lengths are not supported and will be deleted.
ratchIter: Numeric specifying number of iterations of the parsimony ratchet Nixon1999TreeSearch to conduct.
tbrIter: Numeric specifying the maximum number of TBR break points to evaluate before concluding each search. The counter is reset to zero each time tree score improves. The counter is reset to zero each time tree score improves. One "iteration" comprises breaking a single branch and evaluating all possible reconnections.
startIter: Numeric: an initial round of tree search with startIter × tbrIter TBR break points is conducted in order to locate a local optimum before beginning ratchet searches.
finalIter: Numeric: a final round of tree search will evaluate finalIter × tbrIter TBR break points, in order to sample the final optimal neighbourhood more intensely.
maxHits: Numeric specifying the maximum times that an optimal parsimony score may be hit before concluding a ratchet iteration or final search concluded.
maxTime: Numeric: after maxTime minutes, stop tree search at the next opportunity.
quickHits: Numeric: iterations on subsampled datasets will retain quickHits × maxHits trees with the best score.
concavity: Numeric specifying concavity constant for implied step weighting. The most appropriate value will depend on the dataset, but values around 10--15 often perform well Goloboff2018,Smith2019TreeSearch. The character string "profile" employs an approximation of profile parsimony Faith2001TreeSearch. Set as Inf for equal step weights, which underperforms step weighting approaches Goloboff2008,Goloboff2018,Goloboff2019,Smith2019TreeSearch.
ratchEW: Logical specifying whether to use equal weighting during ratchet iterations, improving search speed whilst still facilitating escape from local optima.
tolerance: Numeric specifying degree of suboptimality to tolerate before rejecting a tree. The default, sqrt(.Machine$double.eps), retains trees that may be equally parsimonious but for rounding errors. Setting to larger values will include trees suboptimal by up to tolerance in search results, which may improve the accuracy of the consensus tree (at the expense of resolution) Smith2019TreeSearch.
constraint: Either an object of class phyDat, in which case returned trees will be perfectly compatible with each character in constraint; or a tree of class phylo, all of whose nodes will occur in any output tree. See ImposeConstraint() and vignette for further examples.
verbosity: Integer specifying level of messaging; higher values give more detailed commentary on search progress. Set to 0 to run silently.
method: Unambiguous abbreviation of jackknife or bootstrap specifying how to resample characters. Note that jackknife is considered to give more meaningful results.
proportion: Numeric between 0 and 1 specifying what proportion of characters to retain under jackknife resampling.
...: Additional parameters to MaximizeParsimony().

Resampling

Note that bootstrap support is a measure of the amount of data supporting a split, rather than the amount of confidence that should be afforded the grouping. "Bootstrap support of 100% is not enough, the tree must also be correct" Phillips2004TreeSearch. See discussion in Egan2006;textualTreeSearch; Wagele2009;textualTreeSearch; Simmons2011TreeSearch; Kumar2012;textualTreeSearch.

For a discussion of suitable search parameters in resampling estimates, see Muller2005;textualTreeSearch. The user should decide whether to start each resampling from the optimal tree (which may be quicker, but result in overestimated support values as searches get stuck in local optima close to the optimal tree) or a random tree (which may take longer as more rearrangements are necessary to find an optimal tree on each iteration).

For other ways to estimate clade concordance, see SiteConcordance().

Author

Martin R. Smith (martin.smith@durham.ac.uk)

Details

Tree search commences with ratchIter iterations of the parsimony ratchet Nixon1999TreeSearch, which bootstraps the input dataset in order to escape local optima. A final round of tree bisection and reconnection (TBR) is conducted to broaden the sampling of trees.

This function can be called using the R command line / terminal, or through the "shiny" graphical user interface app (type EasyTrees() to launch).

For detailed documentation of the "TreeSearch" package, including full instructions for loading phylogenetic data into R and initiating and configuring tree search, see the package documentation.

References

Examples

Run this code

## Only run examples in interactive R sessions
if (interactive()) {
  # launch "shiny" point-and-click interface
  EasyTrees()
  
  # Here too, use the "continue search" function to ensure that tree score
  # has stabilized and a global optimum has been found
}


# Load data for analysis in R
library("TreeTools")
data("congreveLamsdellMatrices", package = "TreeSearch")
dataset <- congreveLamsdellMatrices[[42]]

# A very quick run for demonstration purposes
trees <- MaximizeParsimony(dataset, ratchIter = 0, startIter = 0,
                           tbrIter = 1, maxHits = 4, maxTime = 1/100,
                           concavity = 10, verbosity = 4)
names(trees)

# In actual use, be sure to check that the score has converged on a global
# optimum, conducting additional iterations and runs as necessary.
 
if (interactive()) {
# Jackknife resampling
nReplicates <- 10
jackTrees <- replicate(nReplicates,
  #c() ensures that each replicate returns a list of trees
  c(Resample(dataset, trees, ratchIter = 0, tbrIter = 2, startIter = 1,
             maxHits = 5, maxTime = 1 / 10,
             concavity = 10, verbosity = 0))
 )

# In a serious analysis, more replicates would be conducted, and each
# search would undergo more iterations.

# Now we must decide what to do with the multiple optimal trees from
# each replicate.

# Treat each tree equally
JackLabels(ape::consensus(trees), unlist(jackTrees, recursive = FALSE))

# Take the strict consensus of all trees for each replicate
JackLabels(ape::consensus(trees), lapply(jackTrees, ape::consensus))

# Take a single tree from each replicate (the first; order's irrelevant)
JackLabels(ape::consensus(trees), lapply(jackTrees, `[[`, 1))
}

# Tree search with a constraint
constraint <- MatrixToPhyDat(c(a = 1, b = 1, c = 0, d = 0, e = 0, f = 0))
characters <- MatrixToPhyDat(matrix(
  c(0, 1, 1, 1, 0, 0,
    1, 1, 1, 0, 0, 0), ncol = 2,
  dimnames = list(letters[1:6], NULL)))
MaximizeParsimony(characters, constraint = constraint, verbosity = 0)

Run the code above in your browser using DataLab