Tripartitions do not always discriminate phylogenetic networks

Phylogenetic networks are a generalization of phylogenetic trees that allow for the representation of non-treelike evolutionary events, like recombination, hybridization, or lateral gene transfer. In a recent series of papers devoted to the study of reconstructibility of phylogenetic networks, Moret, Nakhleh, Warnow and collaborators introduced the so-called {tripartition metric for phylogenetic networks. In this paper we show that, in fact, this tripartition metric does not satisfy the separation axiom of distances (zero distance means isomorphism, or, in a more relaxed version, zero distance means indistinguishability in some specific sense) in any of the subclasses of phylogenetic networks where it is claimed to do so. We also present a subclass of phylogenetic networks whose members can be singled out by means of their sets of tripartitions (or even clusters), and hence where the latter can be used to define a meaningful metric.

💡 Research Summary

The paper critically examines the “tripartition metric” that has been proposed for comparing phylogenetic networks—graphical models that extend phylogenetic trees to accommodate reticulate evolutionary events such as recombination, hybridization, and horizontal gene transfer. The metric, introduced by Moret, Nakhleh, Warnow and collaborators, assigns to each internal node a partition of the leaf set into three subsets (often denoted A, B, and C). The collection of all such tripartitions for a network constitutes its “signature.” The distance between two networks is defined as the size of the symmetric difference between their signatures. In earlier work the authors claimed that, for several important subclasses of phylogenetic networks (e.g., level‑1 networks, normalized binary networks, certain directed acyclic graph (DAG) models), this distance satisfies the separation axiom: distance zero if and only if the networks are isomorphic (or, in a relaxed sense, indistinguishable under a prescribed equivalence).

The present study demonstrates that these claims are incorrect. First, the authors prove that tripartitions carry exactly the same information as the set of clusters (i.e., the subsets of leaves descended from each node). Consequently, any limitation of cluster‑based discrimination automatically applies to tripartitions. Using this observation, they construct explicit counter‑examples: pairs of non‑isomorphic networks that nonetheless generate identical tripartition (and cluster) sets. The examples are carefully designed to belong to each of the subclasses previously claimed to be discriminated by the metric. For instance, two level‑1 networks are built where the reticulation cycles are arranged differently but each internal node yields the same three‑way leaf split. Similar constructions are provided for level‑2 and higher‑level networks, as well as for unrestricted DAG representations. In every case the tripartition distance evaluates to zero despite the networks being structurally distinct, thereby violating the separation axiom.

Having shown that the metric fails in general, the authors turn to identifying a restricted class of networks for which tripartitions do uniquely determine the structure. They define a “tripartition‑distinguishable class” characterized by three stringent conditions: (1) every internal node has exactly two children (binary tree topology), (2) the network contains no cycles (i.e., it is a tree‑shaped DAG), and (3) each leaf induces a unique cluster, with the inclusion relations among clusters forming a tree. Within this class, the set of tripartitions (or equivalently, the set of clusters) is shown to be a complete invariant: two networks share the same tripartition set if and only if they are isomorphic. Formal theorems and proofs are provided to substantiate this claim.

The discussion emphasizes the practical implications of these findings. Researchers who have employed the tripartition metric for network reconstruction, comparison, or clustering must first verify that their data satisfy the restrictive conditions of the tripartition‑distinguishable class; otherwise, the metric may conflate distinct evolutionary histories. Conversely, for networks that do meet the criteria, the metric remains a valuable tool because it is computationally simple and compatible with existing distance‑based algorithms. The authors suggest several avenues for future work: (i) developing stronger invariants that combine tripartitions with additional graph‑theoretic features (e.g., reticulation edge patterns, path‑frequency vectors), (ii) designing hybrid distances that retain computational efficiency while guaranteeing separation for broader network families, and (iii) establishing empirical validation pipelines using simulated and real biological datasets to assess the robustness of any proposed metric.

In conclusion, the paper provides a rigorous refutation of the previously asserted universal discriminative power of the tripartition metric, clarifies the exact circumstances under which it can be safely applied, and outlines a research agenda aimed at constructing more reliable distance measures for the increasingly complex phylogenetic networks encountered in modern evolutionary biology.