Topology Comparison Study

A topology comparison answers a structural question about two identified trees. It does not compare likelihoods, branch-length accuracy, support calibration, or biological plausibility unless those are separately measured.

Write the question before computing the distance: “Are these two trees equal with respect to <rooted or unrooted splits> over <admitted taxa>?” If the question cannot name both dimensions, a Robinson–Foulds value has no stable interpretation.

Read Robinson–Foulds As A Split-Set Observation

For the admitted taxon set and selected rooting mode, the raw RF distance counts splits present in one tree but not the other. The normalized value scales that difference by the eligible split population used by the implementation. Neither value is a percentage of taxa that “match.”

Change between comparisons	Why the RF value is no longer directly comparable
taxon pruning or addition	the eligible split sets and maximum distance change
rooted versus unrooted mode	directed clades and unrooted bipartitions answer different questions
resolving or collapsing a polytomy	the set of represented splits changes
duplicate or normalized labels	taxon correspondence and therefore split identity changes

Report raw distance, normalized distance, eligible split denominator, and admitted taxa together. A smaller normalized value after additional pruning can reflect an easier comparison population rather than greater agreement between the original trees.

Define The Comparison Population

Record both tree identities, rootedness, taxon sets, duplicate-label checks, and the overlap policy before calculating a distance. The default public workflow prunes to shared taxa; therefore its denominator may be smaller than either input tree.

from pathlib import Path

from bijux_phylogenetics.api import run_tree_comparison_workflow

comparison = run_tree_comparison_workflow(
    Path("study/candidate.nwk"),
    Path("study/reference.nwk"),
    rf_mode="rooted",
    taxon_overlap_policy="prune-to-shared",
)

comparison.write_json(Path("artifacts/topology/comparison.json"))
comparison.write_tsv(Path("artifacts/topology/comparison.tsv"))

The result records the input paths, shared taxa, rooted and unrooted Robinson–Foulds distances, normalized distance, and topology-equality flags. Persist the JSON when the full admitted population matters; the TSV is a compact projection for tabular review.

Preserve an admission ledger beside that result:

Item	Required record
left and right trees	content identity, source, parsing state, and rooting interpretation
label integrity	duplicate, empty, normalized, and unmatched labels
overlap	taxa unique to each tree and the exact shared set
comparison mode	rooted or unrooted split semantics and pruning policy
denominator	eligible split count used by the raw and normalized distances

Without the unique-taxon lists and eligible-split denominator, a normalized distance cannot be audited or compared with another run.

Interpret Distance In Context

flowchart LR
    inputs["Two identified trees"] --> admission["Rooting and shared-taxon<br/>admission"]
    admission --> splits["Comparable split sets"]
    splits --> distance["RF observations"]
    distance --> claim["Bounded topology claim"]

Observation	What it can support	What it cannot support alone
topology_equal	equality under the selected rooted comparison	equal branch lengths, support, likelihood or provenance
same_unrooted_topology	equality after ignoring root placement	agreement about the root
normalized RF	structural difference relative to the eligible split count	comparability across changed taxon populations without qualification
shared taxon count	the admitted comparison population	completeness of either original tree

A zero distance after pruning proves equality only for the retained taxa under the selected mode. List taxa unique to either input and do not summarize that result as whole-tree identity.

Separate Support From Topology

Bootstrap or posterior support labels annotate clades; they are not part of a plain RF topology comparison. Compare support with a clade-keyed ledger that defines absent clades, incompatible splits, and missing values. Do not align support rows by node number or traversal order across independently parsed trees.

Review A Collection Of Trees

For multiple candidates, declare the complete candidate set before selection. Retain one comparison record per eligible pair, failures and excluded trees, the ranking or acceptance rule, and ties. Reporting only the nearest candidate changes the denominator and prevents review of search or sensitivity coverage.

Accept a topology claim only when input identity, rooting, taxon admission, comparison mode, eligible split denominator, and result all remain linked. Create a new comparison record whenever any of those facts changes.

State the accepted conclusion at the same resolution as the computation. For example: “the trees have the same unrooted topology over the 68 shared taxa” is reviewable; “the analyses agree” silently adds branch lengths, support, sampling, and method quality that RF distance did not inspect.