Changelog#

Version 2.0.1#

Bug fixes

  • Reconstructed networks no longer carry placeholder inheritance probabilities. Reticulation edges were previously assigned a hardcoded gamma=0.5, which was written to eNewick output even though Squirrel does not estimate gammas.

Other

  • New physquirrel logo in the documentation and README.

  • Fixed the license badge in the README.

Version 2.0.0#

physquirrel 2.0 is a major release with a new architecture, new data types, and several algorithmic improvements.

Breaking changes

  • All core datatypes (phylogenetic networks, splits, quartets, distance matrices, MSAs) are now provided by phylozoo. The previous self-contained implementations have been removed.

  • QuarnetSet class removed; replaced by SqQuartetProfileSet.

  • Runtime dependencies changed: matplotlib and scipy removed; phylozoo (≥ 0.1.2) and numba (≥ 0.56) added.

  • The [viz] optional dependency now relies on PhyloZoo’s plotting utilities (physquirrel[viz]).

New data types

  • SqQuartetProfile — extends PhyloZoo’s QuartetProfile; holds 1–2 resolved quartets and an optional reticulation_leaf for 4-cycle profiles.

  • SqQuartetProfileSet — collection of SqQuartetProfile objects, one per 4-leaf subset, with optional per-profile weights.

New I/O formats

  • .psq (JSON) — lossless, compact serialization via to_psq() / from_psq().

  • Profile list (plain text) — human-readable SQ / 4C format via to_profile_list() / from_profile_list().

New algorithms

  • quartet_distance_with_partition() — weighted quartet distance between a profile set and a partition, used internally in cycle resolution.

  • split_support() — refactored to correctly handle cycle profiles when scoring candidate splits.

Algorithmic improvements

  • New representative_mode parameter in squirrel() ('best' or 'average'):

    • 'best' (default) — elects a plurality quarnet per 4-tuple of partition sets, matching v1 behaviour.

    • 'average' — averages rho-distances across all representative leaf-partitions for a fuller use of the empirical signal.

  • Weighted profile support: profiles can carry per-profile confidence weights that propagate into cycle resolution and scoring.

  • Parallel execution via ParallelConfig; supports both threading and multiprocessing backends.

  • JIT-accelerated Held-Karp TSP solver via numba (replaces scipy).

Other

  • Full Sphinx documentation with API reference, manual, quickstart, and installation guide.

  • Comprehensive test suite: 242 tests including integration tests on HIV-1, alignments.

Version 1.0.7#

Last release of the v1.x series.