Beyond representing orthology relations by trees

Reconstructing the evolutionary past of a family of genes is an important aspect of many genomic studies. To help with this, simple operations on a set of sequences called orthology relations may be employed. In addition to being interesting from a practical point of view they are also attractive from a theoretical perspective in that e. g. a characterization is known for when such a relation is representable by a certain type of phylogenetic tree. For an orthology relation inferred from real biological data it is however generally too much to hope for that it satisfies that characterization. Rather than trying to correct the data in some way or another which has its own drawbacks, as an alternative, we propose to represent an orthology relation $\delta$ in terms of a structure more general than a phylogenetic tree called a phylogenetic network. To compute such a network in the form of a level-1 representation for $\delta$, we introduce the novel {\sc Network-Popping} algorithm which has several attractive properties. In addition, we characterize orthology relations $\delta$ on some set $X$ that have a level-1 representation in terms of eight natural properties for $\delta$ as well as in terms for level-1 representations of orthology relations on certain subsets of $X$.


Introduction
Unraveling the evolutionary past of a family G of genes is an important aspect for many genomic studies. For this, it is generally assumed that the genes in G are orthologs, that is, have arisen from a common ancestor through speciation. However it is known that shared ancestry of genes can also arise via whole genome duplication (paralogs). This potentially obscures the signal used for reconstructing the evolutionary past of the genes in G in the form of a gene tree (essentially a rooted tree whose leaves are labelled by the elements of G -we present precise definitions of the main concepts used in the next section). To tackle this problem, tree-based approaches have been proposed. These typically work by reconciling a gene tree with an assumed further tree (species tree) in terms of a map that operates on their vertex sets. For this, certain evolutionary events are postulated such as the ones mentioned above (see e. g. [12] for a recent review as well as e. g. [10] and the references therein).
To overcome the problem that the resulting reconciliation very much depends on the quality of the employed trees and also that such approaches can be computationally demanding for larger datasets, orthology relations have been proposed as an alternative. These operate directly on the set of sequences from which a gene tree is built (see e. g. [1]). In addition to having attractive practical properties, such relations are also interesting from a theoretical point of view due to their relationship with e. g. co-trees (see e. g. [5,6]). Furthermore, a characterization is known for when an orthology relation can be represented in terms of a certain type of phylogenetic tree [5].
Due to e. g. errors or noise in an orthology relation, it is however in general too much to hope for that an orthology relation obtained from a real biological dataset satisfies that characterization. A natural strategy therefore might be to try and correct for this in some way. As was pointed out in [11] however, even if an underlying tree-like evolutionary scenario is assumed for this many natural formalizations lead to NP-complete problems. Furthermore, true non-treelike evolutionary signal such as hybridization might be overlooked. As an alternative, we propose to represent orthology relations in terms of phylogenetic networks. These naturally generalize phylogenetic trees by permitting additional edges. To infer such a structure from an orthology relation δ, we introduce the novel Network-Popping algorithm which returns a level-1 representation of δ in In all three cases the underlying phylogenetic network is a level-1 network. Furthermore, N 2 , is not semi-discriminating but weakly labelled whereas N 3 is semidiscriminating but not weakly labelled -See text for details.
the form of a structurally very simple phylogenetic network called a level-1 network (see e. g. Fig. 1 for examples of such representations where the interior vertices labelled in terms of • and • represent two distinct evolutionary events such as speciation and whole genome duplication and the unlabelled interior vertices represent hybridization events). Bearing in mind the point made in [3,Chapter 12], that k-estimates, k ≥ 3, are potentially more accurate than mere distances as they capture more information, we formalize an orthology relation in terms of a symbolic 3-dissimilarity rather than a symbolic 2-dissimilarity (i.e. a distance), as was the case in [5]. From a technical point of view this also allows us to overcome the problem that using a symbolic 2-dissimilarity in a network context can be problematic. An example illustrating this is furnished by the three level-1 representations depicted in Fig. 1 which all represent the same 2-dissimilarity induced by taking the lowest common ancestor between pairs of leaves.
As we shall see, algorithm Network-Popping is guaranteed to find, in polynomial time, a level-1 representation of a symbolic 3-dissimilarity if such a representation exists. For this, it relies on the three further algorithms below which we also introduce. It works by first finding for a symbolic 3-dissimilarity δ on X all pairs of subsets of X that support a cycle using algorithm Find-Cycles. Subsequent to this, it employs algorithm Build-Cycles to construct from each such pair (H, R ) a structurally very simple level-1 representation for the symbolic 3-dissimilarity induced on H ∪ R . Combined with algorithm Vertex-Growing which constructs a symbolic discriminating representation for a symbolic 2-dissimilarity, Network-Popping then recursively grows the level-1 representation for δ by repeatedly applying algorithms Build-Cycles and Vertex-Growing in concert. For the convenience of the reader, we illustrate all four algorithms by means of the level-1 representations depicted in Fig. 1. As part of our analysis of algorithm Network-Popping, we characterize level-1 representable symbolic 3-dissimilarities δ on X in terms of eight natural properties (P1) -(P8) enjoyed by δ. (Theorem 7.1). Furthermore, we characterize such dissimilarities in terms of level-1 representable symbolic 3-dissimilarities on subsets of X of size |X| − 1 (Theorem 8.3). Within a Divide-and-Conquer framework the resulting speed-up of algorithm Network-Popping might allow it to also be applicable to large datasets.
The paper is organized as follows. In the next section, we present basic definitions and results. Subsequent to this, we introduce in Section 3 the crucial concept of a δ-trinet associated to a symbolic 3-dissimilarity and state Property (P1). In Section 4, we present algorithm Find-Cycles as well as Properties (P2) and (P3). In Section 5, we introduce and analyse algorithm Build-Cycles. Furthermore, we state Properties (P4) -(P6). In Section 6, we present algorithms Vertex-Growing and Network-Popping. As suggested by the example in Fig. 1, algorithm Network-Popping need not return the level-1 representation of a symbolic 3-dissimilarity that induced it. Employing a further algorithm called Transform, we address in Section 7 the associated uniqueness question (Corollary 7.5). As part of this we establish Theorem 7.1 which includes stating Properties (P7) and (P8). In Section 8, we establish Theorem 8.3. We conclude with Section 9 where we present research directions that might be worth pursuing.

Basic definitions and results
In this section, we collect relevant basic terminology and results concerning phylogenetic networks and symbolic 2-and 3-dissimilarities. From now on and unless stated otherwise, X denotes a finite set of size n ≥ 3, M denotes a finite set of symbols of size at least two and denotes a symbol not already contained in M . Also, all directed/undirected graphs have no loops or multiple directed/undirected edges.

Directed acyclic graphs.
Suppose G is a rooted directed acyclic graph (DAG), that is, a DAG with a unique vertex with indegree zero. We call that vertex the root of G, denoted by ρ G . Also, we call the graph U (G) obtained from G by ignoring the directions of its edges the underlying graph of G. By abuse of terminology, we call an we say that a vertex v is a leaf if the indegree of v is one and its outdegree is zero. We denote the set of interior vertices of G by V (G) int and the set of leaves of G by L(G). We call a vertex v of G a tree vertex if the indegree of v is at most one and its outdegree is at least two, and a hybrid vertex of G if the indegree of v is two and its outdegree is not zero. The set of interior vertices of G that are not hybrid vertices of G is denoted by V (G) − int . We say that N is binary if, with the exception of ρ N , the indegree and outdegree of each of its interior vertices add up to three. Finally, we say that two DAG's N and N with leaf set X are isomorphic if there exists a bijection from V (N ) to V (N ) that extends to a (directed) graph isomorphism between N and N which is the identity on X.

Phylogenetic networks and last common ancestors.
A (rooted) phylogenetic network N (on X) is a rooted DAG that does not contain a vertex that has indegree and outdegree one and L(N ) = X. In the special case that a phylogenetic network N is such that each of its interior vertices belongs to at most one cycle we call N a a level-1 (phylogenetic) network (on X). Note that a phylogenetic network may contain cycles of length three and that a phylogenetic network that does not contain a cycle is called a phylogenetic tree T (on X).
For the following, let N denote a level-1 network on X. For Y ⊆ X with |Y | ≥ 3, we denote by N | Y the subDAG of N induced by Y (suppressing any resulting vertex that have indegree and outdegree one). Clearly, N | Y is a phylogenetic network on Y .
Suppose v is a non-leaf vertex of N . We say that a further vertex w ∈ V (N ) is below v if there is a directed path from v to w and call the set of leaves of N below v the offspring set of v, denoted by F(v). Note that F(v) is closely related to the hardwired cluster of N induced by v (see e.g. [9]). For a leaf x ∈ F(v), we refer to v as an ancestor of x. In case N is a phylogenetic tree, we define the lowest common ancestor lca N (x, y) of two distinct leaves x, y ∈ L(N ) to be the (necessarily unique) Note that in case the tree N we are referring too is clear from the context, we shall write lca(Y ) rather than lca N (Y ).
It is easy to see that the notion of a lowest common ancestor is not well-defined for phylogenetic networks in general. However the situation changes in case the network in question is a level-1 network, as the following central result shows. Since its proof is straight-forward, we omit it.
Continuing with the terminology of Lemma 2.1, we shall refer to v Y as the lowest common ancestor of Y in N , denoted by lca N (Y ). As in the case of a phylogenetic tree, we shall write lca(Y ) rather than lca N (Y ) if the network N we are referring to is clear from the context. 2.3. Symbolic dissimilarities and labelled level-1 networks. Suppose k ∈ {2, 3}. We denote by X k the set of subsets of X of size k, and by X ≤k the set of nonempty subsets of X of size at most k. We call a map δ : ≤k , we have that δ(A) = if and only if |A| = 1. To improve clarity of exposition, we shall refer to δ as a symbolic 3-dissimilarity on X if the set M is of no relevance to the discussion. Moreover, for Y = {x 1 , . . . , x l }, l ≥ 2, we shall write δ(x 1 , . . . , x l ) rather than δ(Y ) where the order of the elements x i , 1 ≤ i ≤ l, is of no relevance to the discussion.
A labelled (phylogenetic) network N = (N, t) (on X) is a pair consisting of a phylogenetic network N on X and a labelling map t : V (N ) − int → M . If N is a level-1 network then N is called a labelled level-1 network. To improve clarity of exposition we shall always use calligraphic font to denote a labelled phylogenetic network.
Suppose Conversely, suppose δ is a symbolic 3-dissimilarity on X. In view of Lemma 2.1, we call a labelled level-1 network N = (N, t) on X a level-1 representation of δ if δ = δ N . For ease of terminology, we shall sometimes say that δ is level-1 representable if the the labelled network we are referring too is of no relevance to the discussion. We call a level-1 representation of δ semi-discriminating if N does not contain a directed edge (u, v) such that t(u) = t(v) except for when there exists a cycle C of N with |V (C) ∩ {u, v}| = 1. For example, all three labelled level-1 networks depicted in Fig. 1 are level-1 representations of δ N 2 where N 2 is the labelled level-1 network depicted in Fig. 1(ii). Furthermore, the representations of δ N 2 presented in Fig. 1(i) and (iii), respectively, are semi-discriminating whereas the one depicted in Fig. 1(ii) is not.
Note that in case N is a phylogenetic tree on X the definition of a semi-discriminating level-1 representation for δ reduces to that of a discriminating symbolic representation for the restriction δ 2 = δ| ( X ≤2 ) of δ to X 2 (see [2] and also [5,13] for more on such representations). Using the concept of a symbolic ultrametric, that is, a symbolic 2dissimilarity δ : X ≤2 → M for which, in addition, the following two properties are satisfied (U1) |{δ(x, y), δ(x, z), δ(y, z)}| ≤ 2 for all x, y, z ∈ X; (U2) there exists no four elements x, y, z, u ∈ X such that Clearly, it is too much to hope for that any symbolic 3-dissimilarity δ has a level-1 representation. The question therefore becomes: Which symbolic 3-dissimilarities have such a representation? A first partial answer is provided by Theorem 2.2 and Lemma 2.1 for not δ but its restriction δ 2 . More precisely, δ has a discriminating symbolic representation if and only if δ 2 is a symbolic ultrametric and, for all x, y, z ∈ X distinct, δ(x, y, z) is the (unique) element appearing at least twice in the multiset {δ 2 (x, y), δ 2 (x, z), δ 2 (y, z)}.

δ-triplets, δ-tricycles, and δ-forks
To make a first inroad into the aforementioned question, we next investigate structurally very simple level-1 representations of symbolic 3-dissimilarities. As we shall see, these will turn out to be of fundamental importance for our algorithm Network-Popping (see Section 6) as well as for our analysis of its properties. In the context of this, it is important to note that although triplets (i. e. binary phylogenetic trees on 3 leaves) are well-known to uniquely determine (up to isomorphism) phylogenetic trees this does not hold for level-1 networks in general [4]. To overcome this problem, trinets, that is, phylogenetic networks on three leaves were introduced in [7]. For the convenience of the reader, we depict in Fig. 4 all 12 trinets τ 1 , . . . , τ 12 on X = {x, y, z} from [7] that are also level-1 networks in our sense. In the same paper it was observed that even the slightly more general 1-nested networks are uniquely determined by their induced trinet sets (see also [8] for more on constructing level-1 networks from trinets, and [14] for an extension of this result to other classes of phylogenetic networks).
Perhaps not surprisingly, trinets on their own are not strong enough to uniquely determine labelled level-1 networks in the sense that any two level-1 representations of a symbolic 3-dissimilarity must be isomorphic. To see this, suppose |X| = 3 and consider the symbolic 3-dissimilarity δ : X ≤3 → {A, B, } that maps X and every 2-subset of X to A. Then the labelled network (τ 1 , t) where t maps every vertex in V (τ 1 ) − int to A is a semi-discriminating level-1 representation of δ and so is the labelled network (τ 4 , t ), where every vertex in V (τ 4 ) − int is mapped to A by t . Note that similar arguments may τ12 : Figure 4. The twelve trinets in the from of level-1 networks. The two omitted trinets from [7] are not level-1 networks in our sense. also be applied to the level-1 representations involving the trinet τ 4 to τ 12 depicted in Fig. 4. We therefore evoke parsimony and focus for the remainder of this paper on the trinets τ 1 , τ 2 and τ 3 . We shall refer to them as fork on X = {x, y, z}, triplet z|xy, and tricycle y||xz, respectively. The next result (Lemma 3.1) relates forks, triplets and tricycles with symbolic 3dissimilarities. To state it, we say that a symbolic 3-dissimilarity δ satisfies the Helly-type Property if, for any three elements x, y, z ∈ X, we have δ(x, y, z) ∈ {δ(x, y), δ(x, z), δ(y, z)}. Note that we will sometimes also refer to the Helly-type property as Property (P1). Proof. Suppose first that N = (N, t) is a level-1 representation of δ. Then, in view of Conversely, suppose that δ(x, y, z) ∈ E := {δ(x, y), δ(x, z), δ(y, z)} holds for all elements x, y, z ∈ X distinct. By analyzing the size of E it is straight-forward to show that one of the situations indicated in the rightmost column of Table 1 must apply. With defining a labelling map t : V (N ) − int → M in the obvious way using the second column of that table, it follows that N is a level-1 representation for δ.
Armed with Lemma 3.1, we make the following central definition. Suppose that |Y | = 3, that δ is a symbolic 3-dissimilarity on Y , and that N = (N, t) is a semidiscriminating level-1 representation of δ. Then we call N a δ-fork if N is a fork on Y , a δ-triplet if N is a triplet on Y , and a δ-tricycle if N is a tricycle on Y , For ease of terminology, we will collectively refer to all three of them as a δ-trinet. Note that as the example of the labelled trinet depicted in Fig. 3(iii) shows, there exist trinets that are Table 1. For δ : X ≤3 → M a symbolic 3-dissimilarity we list all labelled trinets on X = {x, y, z} in terms of the size of E.
not δ-trinets. By abuse of terminology, we shall refer for a symbolic 3-dissimilarity δ on X and any 3-subset Y ⊆ X to a δ| Y -trinet as a δ-trinet.

Recognizing cycles: The algorithm Find-Cycles
In this section, we introduce and analyze algorithm Find-Cycles (see Algorithm 1 for a pseudo-code version). Its purpose is to recognize cycles in a level-1 representation of a symbolic 3-dissimilarity δ if such a representation exists. As we shall see, this algorithm relies on Property (P1) and a certain graph C(δ) that can be canonically associated to δ. Along the way, we also establish two further crucial properties enjoyed by a level-1 representable symbolic 3-dissimilarity.
We start with introducing further terminology. Suppose N is a level-1 network and C is a cycle of N . Then we denote by r(C) the unique vertex in C for which both children are also contained in C and call it the root of C. In addition, we call the hybrid vertex of N contained in C the hybrid of C and denote it by h(C). Furthermore, we denote set of all elements of X below by r(C) by R(C) and the set of all elements of X below h(C) by H(C). Clearly, H(C) R(C). Moreover, for any leaf x ∈ R(C) − H(C), we denote by v C (x) the last ancestor of x in C. Note that v C (x) is the parent of x if and only if x is incident with a vertex in C. Last-but-not-least, we call the vertex sets of the two edge-disjoint directed paths from r(C) to h(C) the sides of C. Denoting these two paths by P 1 and P 2 , respectively, we say that two leaves x and y in R(C) − H(C) lie on the same side of C if the vertices v C (x) and v C (y) are both interior vertices of P 1 or P 2 , and that they lie on different sides if they are not. For example, for C the underlying cycle of the cycle C 2 indicated in the labelled network N 1 pictured in Fig. 1 , h(C)} and d, . . . , g lie on one side of C whereas b and d lie on different sides of C.
Suggested by Property (U2), the following property is of interest to us where δ denotes again a symbolic 3-dissimilarity on X: such that a tricycle on Y underlies a level-1 representation of δ| Y . As a first result, we obtain Proof. Note first that Property (P1) is a straight-forward consequence of Lemma 2.1.
To see that Property (P2) holds, note first that since δ is level-1 representable there exists a labelled level-1 network ( is a δ-tricycle, assume for contradiction that there exists no such set Y . By Theorem 2.2, N cannot be a phylogenetic tree on X and, so, N must contain at least one cycle C. Without loss of generality, we may assume that x ∈ H(C), and y lies on one of the two sides of C. By assumption δ(y, z) = δ(x, z) and so either z and y lie on opposite sides of C, or z and y lie on the same side of C and v C (y) lies on the directed path from r(C) to v C (z). As can be easily checked, either one of these two cases yields a contradiction since then δ(z, u) = δ(x, u) = δ(y, u) cannot hold for u, as required.
To see that there can exist at most one such tricycle on Z, assume for contradiction that there exist tow tricycles τ and τ with . Note that the assumption on the elements of Z implies that x or y must be below the hybrid vertex of one of τ and τ but not the other. Without loss of generality we may assume that y is below the hybrid vertex of τ but not below the hybrid vertex of τ . Then y must lie on a side of the unique cycle C of τ . But this is impossible since the unique cycle of τ and C are induced by the same cycle of N .
We remark in passing that the proof of uniqueness in the proof of Lemma 4.1 combined with the structure of a level-1 network, readily implies the following result. (N, t) and that x, y, z ∈ X are three distinct elements such that x||yz is a δ-tricycle. Let C denote the unique cycle in N such that x ∈ H(C) and y, z ∈ R(C) − H(C), and let x ∈ X. If x ||yz is a δ-tricycle then x ∈ H(C) and if x||x z is a δ-tricycle then x ∈ R(C) and x and y lie on the same side of C.

Lemma 4.2. Suppose that δ is a symbolic 3-dissimilarity on X that is level-1 representable by a labelled network
To better understand the structure of a symbolic 3-dissimilarity δ, we next associate to δ a graph C(δ) defined as follows. The vertices of C(δ) are the δ-tricycles and any two δ-tricycles τ and τ are joined by an edge if |L(τ ) ∩ L(τ )| = 2. For example, consider the symbolic 3-dissimilarity δ N 1 induced by the labelled level-1 network N 1 pictured in Fig. 1(i). Then the graph presented in Fig. 5   . The graph C(δ N 1 ), where N 1 is the labelled level-1 network depicted in Fig. 1(i).
We collect first results concerning Property (P3) in the next proposition.
If δ is level-1 representable or |M | = 2 holds then Property (P3) must hold. In particular, if N is a level-1 representation for δ then there exists a canonical injective map from the set of connected components of C(δ) to the set of cycles of the level-1 network underlying N .
) holds for all cycles C of N , and any x ∈ H(C) and any y, z ∈ R(C) that lie on different sides of C, Property (P3) follows.
The claimed injective map is a straight-forward consequence of Lemma 4.2.
Algorithm Find-Cycles exploits the injection mentioned in Proposition 4.3 by interpreting for a symbolic 3-dissimilarity δ a connected component C of C(δ) in terms of two sets H C and R C . Note that if C is a cycle in the level-1 network underlying a level-1 representation of δ (if such a representation exists!), the sets H(C ) and H C coincide and R C ⊆ R(C ) holds.

Constructing cycles: The algorithm Build-Cycles
We next turn our attention toward reconstructing a structurally very simple level-1 representation of a symbolic 3-dissimilarity (should such a representation exist). For this, we use algorithm Build-Cycles which takes as input a symbolic 3-dissimilarity δ and a pair returned by Find-Cycles when given δ.
To state Build-Cycles, we require further terminology. Suppose N is a level-1 network. Then we say that N is partially resolved if all vertices in a cycle of N have degree three. Note that partially-resolved level-1 networks may have interior vertices not contained in a cycle that have degree three or more. Thus such networks need not be binary. If, in addition to being partially resolved, N is such that it contains a unique cycle C such that every non-leaf vertex of N is a vertex of C then we call N simple.
Algorithm Build-Cycle (see Algorithm 2 for a pseudo-code version) relies on a further graph called the TopDown graph associated to a symbolic 3-dissimilarity δ. For Input: A symbolic 3-dissimilarity δ on X. Output: A number m ≥ 1 and pairs of subsets Build the graph C(δ);

3
Denote by m the number of connected components of C(δ); Proof. Put δ = δ N . Suppose first that v C (z) lies on the directed path from v C (y) to h(C). Then lca(x, y, z) = lca(x, y) = lca(y, z) = v C (y) and lca(x, z) = v C (z). Hence, δ(x, y, z) = δ(x, y) = δ(y, z) = t(v C (y)) = t(v C (z)) = δ(x, z). By Table 1, y|xz is a δ-triplet. Conversely, suppose that y|xz is a δ-triplet. Then, by Table 1, we have δ(x, y, z) = δ(x, y) = δ(y, z) = δ(x, z). Since δ(x, y) = t(v C (y)) and δ(x, z) = t(v C (z)), it follows that δ(x, y, z) = t(v C (y)) = t(v C (z)). But then y and z must lie on the same side of C as otherwise δ(y, z) = t(r(C)) follows which is impossible by assumption on x, y and z. Thus, either v C (y) must lie on a directed path P from v C (z) to h(C) or v C (z) must lie on a directed path P from v C (y) to h(C). However v C (y) cannot be a vertex on P as otherwise lca(y, z) = v C (z) holds and, so, δ(y, z) = δ(x, z) follows, which is impossible.
Thus v C (z) must be a vertex on P .
With N and C as in from Lemma 5.1, it follows from Lemma 4.2, that whenever algorithm Find-Cycles is given δ N as input, it returns a pair (H, R ) such that H = H(C) and R = H(C) ∪ {y ∈ R(C) : t(v C (y)) = t(r(C))}. Moreover giving (H, R ) and δ N as input to algorithm Build-Cycle, Lemma 5.1 implies that Build-Cycle finds all elements z ∈ R(C) − R for which there exists some y ∈ R such that v C (z) lies on the path from v C (y) to h(C). However it should be noted that if z ∈ R(C) − H(C) is such that t(v) = t(r(C)) = t(v C (z)) holds for all vertices v on the path from r(C) to v C (z) then the information captured by δ N for x, y, and z is in general not sufficient to decide if z and y lie on the same side of C or not. In fact, it is easy to see that, in general, z ∈ R(C) need not even hold.
We now turn out attention to the aforementioned TopDown graph associated to a symbolic 3-dissimilarity δ on X which is defined as follows. Suppose that S X, and that x ∈ X − S. Then the vertex set of the TopDown graph T D(S, x) is S and two elements u, v ∈ S distinct are joined by a direct edge (u, v) if u|vx is a δ-triplet.  Fig. 6(a). It implies that for the cycle C associated to the pair (c, bcef g) in a level-1 representation of δ N 1 , we must have v C (e) = v C (f ) = v C (g) and that one of the two sides of C is {e, d, f, g}. Since |S b | = 1, the other side of C is {b} (lines 11 to 33).
Continuing with our analysis of algorithm Build-Cycle, we remark that the fact that the TopDown graph T D(S e , c) in the previous example is non-empty is not a coincidence. In fact, it is easy to see that the graph G defined in line 14 of Build-Cycle is non-empty whenever δ is level-1 representable. Thus, the DAG C returned by algorithm Build-Cycle cannot contain multi-arcs. Note however that there might be tricycles induced by C of the form x||uz with u ∈ R − S y as, for example, δ(x, z) = δ(x, y) = δ(z, y) = δ(x, u) might hold and thus x||uz is not a δ-tricycle. Note that similar reasoning also applies to S z and the extensions of S y and S z to S y and S z defined in lines 8 and 9, respectively. Also note that the sets S y and S z are dependent on the choice of the δ-tricycle in line 2.

Input:
A symbolic 3-dissimilarity δ on X that satisfies Property (P1) and a pair (H, R ) returned by algorithm Find-Cycle when given δ. Output: Either a labelled simple level-1 network (C, t) on a partition of a subset X of X such that R ⊆ X and H(K) = H holds for the unique cycle K of C, or the statement "δ is not level-1 representable".
x||yu is δ-tricycle}; 5 Initialize C as a graph with three vertices respectively labelled by r(C), h(C) and H, and the edge (h(C), H); 6 if for all x ∈ H, y ∈ S y and z ∈ S z , x ||y z is a δ-tricycle and δ(x, y, z) = δ(x , y , z ) then However, line 6 ensures that the labelled simple level-1 network returned by algorithm Build-Cycle is independent of the choice of that δ-tricycle.
To establish Proposition 5.3 which ensures that algorithm Build-Cycle terminates, we next associate to a directed graph G a new graph P (G) by successively removing vertices of indegree zero and their incident edges until no such vertices remain. As a first almost trivial observation concerning that graph we have the following straightforward result whose proof we again omit.

Lemma 5.2. Let G be a directed graph. Then P (G) is nonempty if and only if G contains a directed cycle.
Given as input to algorithm Build-Cycle a symbolic 3-dissimilarity δ that satisfies Property (P1) and a pair (H, R ) returned by algorithm Build-Cycle for δ we have It is straight-forward to see that when given a level-1 representable symbolic 3dissimilarity δ such that the underlying level-1 network is in fact a simple level-1 network the labelled network returned by algorithm Build-Cycle satisfies the following three additional properties (where we use the notations introduced in algorithm Build-Cycle). ( For all u, v ∈ R := H ∪ S y ∪ S z and all w ∈ X − R, we have δ(u, w) = δ(v, w). (P6) For all u, u ∈ H and i ∈ {y, z}, the graphs T D(S i , u) and T D(S i , u ) are isomorphic and do not contain a directed cycle. Since the quantities on which these properties are based also exist for general symbolic 3-dissimilarities we next study Properties (P4) -(P6) for such dissimilarities. As a first consequence of Property (P4) combined with Properties (P1) and (P2), we obtain a sufficient condition under which the TopDown graph T D(S i , x) considered in algorithm Build-Cycle does not contain a directed cycle (lines 13). For convenience, we employ again the notation used in Algorithm 2.  such that (a, b), (b, c), (c, d) are three directed edges in C. We next distinguish between the cases that |V (C)| ≥ 5 and that |V (C)| = 4.
Suppose |V (C)| ≥ 5. Then since a, c ∈ S y , Lemma 4.2 combined with the minimality of C implies that we either have a δ-fork on {a, c, x} or the δ-triplet ac|x. Hence, δ(x, a) = δ(x, c) holds in either case. Note that similar arguments also imply that δ(x, b) = δ(x, d). Since |V (C)| ≥ 5, the directed edges (a, d) and (d, a) cannot be contained in T D(S y , x) and, using again similar arguments as before, δ(x, a) = δ(x, d) must hold. In combination, we obtain δ(x, a) = δ(x, b) which is impossible in view of (a, b) being an edge in T D(S y , x) and thus δ(x, a) = δ(x, b). Suppose Note that, δ(a, c) ∈ {A, B} must also hold as otherwise |{δ(a, c), δ(a, b), δ(b, c)}| = 3 and so, in view of Table 1 Table 1 with Equation 1, it is straight-forward to see that each of the four detailed combinations of δ(a, c) and δ(b, d) in that table yields a contradiction in view of Property (P2).
(ii) By symmetry, it suffices to assume i = y. Let |M | = 2 and assume for contradiction that T D(S y , x) contains a directed cycle C of size 3. Let s, u, v denote the 3 vertices of C such that (s, u), (u, v) and (v, s) are the three directed edges of C. Then δ(u, x) = δ(s, x) = δ(v, s) = δ(v, x) = δ(u, v) = δ(u, x) must hold. Since |M | = 2, this is impossible.

Constructing level-1 representations from symbolic 3-dissimilarities:
The algorithm Network-Popping In this section, we present algorithm Network-Popping which allows us to decide if a symbolic 3-dissimilarity is level-1 representable or not. If it is, then Network-Popping is guaranteed to find a level-1 representation in polynomial time.
Network-Popping takes as input a symbolic 3-dissimilarity δ on X and employs a top-down approach to recursively construct a semi-discriminating level-1 representation for δ (if such a representation exists). For l a leaf whose label set is of size at least two and constructed in one of the previous steps it essentially works by either replacing l with a labelled simple level-1 network or a labelled phylogenetic tree. To compute those networks algorithms Find-Cycle and Build-Cycle are used and to construct such trees algorithm Vertex-Growing is employed. At the heart of the latter lie Proposition 6.2 and algorithm Bottom-Up introduced in [5]. The latter takes as input a symbolic 2-dissimilarity δ satisfying Properties (U1) and (U2), and builds the unique discriminating symbolic representation T for δ (if it exists).
To be able to state algorithm Vertex-Growing, we require again further terminology. Following e. g. [13], we call a collection H of non-empty subsets of X a hierarchy on X if A ∩ B ∈ {A, B, ∅} holds for any two sets A, B ∈ H. The proof of the following result is straight-forward and thus omitted. Lemma 6.1. Let N be a level-1 network with cycles C 1 , C 2 , . . . , C k , k ≥ 1. Then, Suppose A is a set of non-empty subsets of X. Then we define a relation ∼ (X,A) on X by putting x ∼ (X,A) y if there exists some A ∈ A such that x, y ∈ A, for all x, y ∈ X. Note first that ∼ (X,A) is clearly an equivalence relation whenever A is a hierarchy. In addition, suppose that A is such that the partition X of X induced by ∼ (X,A) has size two or more. If δ : X ≤3 → M is a symbolic 3-dissimilarity such that for any two sets Y, Y ∈ X we have δ(x, y) = δ(x , y ) for all x, x ∈ Y and y, y ∈ Y , then we associate to δ the mapδ given bŷ Note thatδ is clearly well-defined and a symbolic 2-dissimilarity on X . Associating to a level-1 representation N = (N, t) of δ the set R := {R(C) : C is a cycle of N }, we have the following result as an immediate consequence. N is a labelled level-1 network on X and X is the partition of X induced by the relation ∼ (X,R) on X. If |X | ≥ 2 thenδ N is well defined and satisfies Properties (U1) and (U2). In particular,δ N is a symbolic ultrametric on X .

Proposition 6.2. Suppose
Proof. Put N = (N, t) and δ =δ N . Note first that for all x, y ∈ X, Lemma 6.1 implies that there exists some R ∈ R such that x, y ∈ R if and only if there exists R ∈ R := {R ∈ R : R is set-inclusion maximal in R} such that x, y ∈ R . Let T N denote the tree obtained from N by first collapsing for every cycle C of N with R(C) ∈ R all vertices below or equal to r(C) into a vertex and then labelling that vertex by R(C). Put t N := t| V (T N ) . Then (T N , t N ) is clearly a labelled phylogenetic tree on X . Since N is a labelled level-1 network, it follows that (T N , t N ) is a symbolic discriminating representation ofδ N . In view of Theorem 2.2, the proposition follows.
To illustrate algorithm Vertex-Growing consider again the symbolic 3-dissimilarity δ N 1 induced by the labelled level-1 network on X = {a . . . , k} depicted in Fig. 1(i).
} is X itself, since any two leaves of X are in relation with respect to ∼ (X,R) . Thus, the discriminating symbolic representation returned by Bottom-Up is a single leaf.
Armed with the algorithms Find-Cycles, Build-Cycles, and Vertex-Growing, we next present a pseudo-code version of algorithm Network-Popping (Algorithm 4).

Input:
A symbolic 3-dissimilarity δ on a set X, a subset Y ⊆ X, and a hierarchy S of proper subsets of Y . Output: A discriminating symbolic representation on the partition of Y induced by ∼ (Y,S) or the statement "There exists no discriminating symbolic representation". To be able to establish in Proposition 6.4 that algorithm Network-Popping returns a semi-discriminating level-1 representation for a symbolic 3-dissimilarity (if such a representation exists), we require the following technical result. Proposition 6.3. Let δ be a symbolic 3-dissimilarity on X satisfying Property (P1), and assume that Network-Popping returns a labelled level-1 network N on X when given δ as input. Then the restrictions δ| ( X ≤2 ) and δ N | ( X ≤2 ) of δ and δ N to X ≤2 , respectively, coincide if and only if δ and δ N coincide.
Conversely, assume that δ = δ N . Let Z = {a, b, c} ∈ X 3 and put m = δ(Z). Note that since N is clearly a level-1 representation of δ N , Lemma 4.1 implies that δ N also satisfies Property (P1). Further note that, up to permuting the elements in Z, we either have (i) a δ-fork on Z, (ii) a|bc is a δ-triplet, or (iii) a||bc is a δ-tricycle.
If Case (i) holds then δ(a, b) = δ(a, c) = δ(b, c) = m. Since, by assumption, If Case (ii) holds then m = δ(a, b) = δ(a, c) = δ(b, c). Assume for contradiction that δ N (Z) = m. Then, since δ N satisfies Property (P1) it follows that δ N (Z) = δ N (b, c). By Table 1, a||bc must be a δ N -tricycle. Hence, there must exist a cycle C in N such that a ∈ H(C), b and c are contained in R(C) but lie on different sides of C, and t(r(C)) = δ N (Z). Since algorithm Network-Popping completes by returning N it follows that C is constructed in the while-loop starting in line 16 of algorithm Build-Cycle. But then the condition in line 6 of Build-Cycle has to be satisfied which implies that t(r(C)) = δ(Z) in view of line 7 of that algorithm. Hence, m = δ N (Z) = t(r(C)) = δ(Z) = m which is impossible.

Input: A symbolic 3-dissimilarity δ on X.
Output: A semi-discriminating level-1 representation N = (N, t ) of δ, if such a representation exists, or the statement "δ is not level-1 representable". 1 Initialize N as an unique vertex v, labelled by X; If Case (iii) holds then the while-loop initiated in line 16 of algorithm Build-Cycle implies that there must exist a cycle C in N such that t(r(C)) = δ(Z) = m. Since N is returned by algorithm Network-Popping when given δ and N is clearly a level-1 representation for δ N it follows that δ N (Z) = t(r(C)) = m = δ(Z).
As a first result concerning algorithm Network-Popping, we have Proposition 6.4. Suppose δ is a symbolic 3-dissimilarity on X, and Network- Popping  applied to δ returns a labelled level-1 network N . Then δ = δ N . In particular, N is a  level-1 representation for δ. Proof. Put N = (N, t). In view of Proposition 6.3, it suffices to show that δ(a, b) = δ N (a, b) holds for all a, b ∈ X distinct. Let a and b denote two such elements. We distinguish between the cases that either (i) there exists a cycle C of N such that v C (a) = v C (b), or (ii) that no such cycle exists.
Assume first that Case (i) holds. Then a and b lie either on the same side of C, or one of a and b is below the hybrid h(C) of C and the other lies on the side of C, or a and b lie on different sides of C. If a and b lie on the same side of C or one of them is below h(C) then we may assume without loss of generality that there exists a directed path in C from v C (a) to v C (b). Then line 22 of algorithm Build-Cycle implies t(v C (a)) = δ(a, b). δ(a, b), as required.
If a and b lie on different sides of C then x||ab is a δ-tricycle, for x as in line 2 of algorithm Build-Cycle. Since that algorithm completes, it's line 7 implies δ(a, b) = t(r(C)). But then δ N (a, b) = t(r(C)) = δ(a, b), as N is returned by Network-Popping.
For the remainder, assume that Case (ii) holds, that is, there exists no cycle C of N such that v C (a) = v C (b). Consider the vertex v 0 ∈ V (N ) defined as follows: if the path from the root ρ N of N to lca(a, b) does not contain a vertex that is also contained in a cycle of N , then put v 0 = ρ N . Otherwise let v 0 denote the last vertex on a directed path from ρ N to lca(a, b) such that v 0 belongs to a cycle Z of N . Note that v 0 = lca(a, b) holds if lca(a, b) is also contained in Z. Put V = F(v 1 ) where v 1 is the unique child of v 0 not contained in Z, and let V denote the partition of V induced by ∼ (V,Sv 0 ) where for any vertex w ∈ V (N ) the set S w is as defined as in line 12 of algorithm Network-Popping. Let R a , R b ∈ V such that a ∈ R a and b ∈ R b . Then line 5 of Network-Popping impliesδ N (R a , R b ) = δ N (a, b) andδ(R a , R b ) = δ(a, b). Since N is returned by Network-Popping when given δ, line 12 of that algorithm implieŝ δ(R a , R b ) =δ N (R a , R b ). Consequently, δ N (a, b) = δ(a, b) holds in this case too.
We conclude this section with remarking that the runtime of algorithm Network-Popping is polynomial. The reasons for this are that Network-Popping basically works by comparing δ-trinets and δ-values on subsets on X of size two, and that the number of such trinets and subsets is polynomial in the size of X.

Uniqueness of level-1 representations returned by Network-Popping
As is easy to see, there exist symbolic 3-dissimilarities that although they satisfy Properties (P1) -(P6) they are not level-1 representable. The reason for this is that such 3-dissimilarities need not satisfy the assumptions of lines 10 and 20 in algorithm Build-Cycle. A careful analysis of that algorithm suggests however two further properties for a symbolic 3-dissimilarity to be level-1 representable. To state them, we next associate to a symbolic 3-dissimilarity its CheckLabels graph.
Suppose Y 0 , Y 1 , and Y 2 are three pairwise disjoint subsets of X such that for all x, x ∈ Y 0 and all i = 1, 2, the graphs T D(Y i , x) and T D(Y i , x ) are isomorphic (which is motivated by Property (P6)). Then we denote by CL(Y 0 , Y 1 , Y 2 ) the CheckLabels graph associated to δ, Y 0 , Y 1 , and Y 2 defined as follows. The vertex set of CL joined by a directed edge (u, v), and two elements u, v ∈ Y i , i = 1, 2, are joined by a directed edge (u, v) if there exists a direct path from u to v in T D (Y i , x). Finally, to each edge of CL(Y 0 , Y 1 , Y 2 ) with end vertices u and v or directed edge of that graph with tail u and head v, we assign the label δ(u, v). We illustrate the CheckLabels graph in Fig. 6(b) for the network N 1 depicted in Fig. 1(i).
Using the terminology of algorithm Build-Cycle it is straight-forward to observe that the following two properties are implied by Build-Cycle's lines 10 and 20 whenever its input symbolic 3-dissimilarity is level-1 representable: (P7) All undirected edges of CL(H, S y , S z ) have the same label; (P8) For all vertices u of CL(H, S y , S z ), all directed edges in CL(H, S y , S z ) with tail u have the same label. As indicated in Table 2, Properties (P1) -(P8) are independent of each other. Moreover, they allow us to characterize level-1 representable symbolic 3-dissimilarities.
Prop.  Since, in view of the while loop of that algorithm starting at line 7, no two cycles in N can share a vertex it follows that N is in fact a level-1 network. Proposition 6.3 combined with the observation that in none of our four algorithms we have to break a tie implies that N is unique up to isomorphism.
(iii) ⇒ (iv): This is trivial in view of Proposition 6.4.
(iv) ⇒ (v): Suppose algorithm Network-Popping returns a level-1 representation N for δ. To see that N is in fact semi-discriminating, note that algorithms Vertex-Growing and Build-Cycles return a discriminating symbolic representation and a discriminating level-1 representation for its input symbolic 3-dissimilarity, respectively. In combination it follows that N must be semi-discriminating.
As suggested by the two semi-discriminating level-1 representations N 1 and N 3 for δ N 1 depicted in Fig. 1(i) and (ii), the output of algorithm Network Popping when given a level-1 representable symbolic 3-dissimilarity δ need not be the labelled level-1 network that induced δ. To help clarify the relationship between both networks, we require further terminology.
Suppose that (N, t) is a labelled level-1 network. Then we say that a cycle C of N is weakly labelled if there exists at least one vertex v on either side of C such that t(v) = t(r(C)). More generally, we call a labelled level-1 network (N, t) weakly labelled if every cycle of N is weakly labelled. For example, the labelled level-1 network N 2 pictured in Fig. 1(ii) is weakly labelled (but not semi-discriminating) whereas the network N 3 depicted in Fig. 1(iii) is semi-discriminating but not weakly labelled.
As a consequence, we can strengthen Proposition 4.3 to the following characterization. Corollary 7.5. Suppose N is a labelled level-1 network and N is the level-1 representation for δ N returned by algorithm Network-Popping. Then N is isomorphic with the labelled level-1 network returned by algorithm Transform when given N as input. In particular, N and N are isomorphic if and only if N is semi-discriminating, weakly labelled, and partially resolved. Furthermore, if δ is a level-1 representable symbolic 3dissimilarity, then there exists an unique representation of δ that is semi-discriminating, weakly labelled, and partially resolved.

Characterizing level-1 representable symbolic 3-dissimilarities
In this section, we present a characterization of level-1 representable symbolic 3dissimilarities on X in terms of level-1 representable symbolic 3-dissimilarities on subsets Input: A labelled level-1 network N = (N, t) on X. Output: A semi-discriminating, weakly labelled, partially resolved level-1 network N = (N , t ) such that δ N = δ N . 1 set N = N ; 2 while N is not semi-discriminating or not weakly labelled or not partially resolved do 3 Collapse all edges (u, v) satisfying t (u) = t (v) and such that either u and v belong to the same cycle of N or do not belong to a cycle; Combined with the fact that algorithm Network-Popping has polynomial run time, this suggests that Network-Popping might lend itself to studies involving large data sets using a Divide-and-Conquer approach. At the heart of the proof of our characterization lies the following technical lemma which concerns the question under what circumstances the restriction of a level-1 representable symbolic 3-dissimilarity δ on X is itself level-1 representable. Central to its proof is the fact that |X| = 4 since, in general, a symbolic 3-dissimilarity δ on a set X of size 4 need not be level-1 representable but the restriction of δ to any subset of size 3 is level-1 representable. An example for this is furnished by the symbolic 3-dissimilarity δ on X = {x, y, z, u}, given by δ(x, y, z) = δ (y, z, u Using the assumptions and definitions for the elements x, y, and z, and the sets H, S z , and S y made in algorithm Build-Cycle, we have the following result. Proof. By symmetry, we may assume i = y. Suppose there exists a directed path v 0 = u, v 1 , . . . , v k , v k+1 = w, some k ≥ 0, from u to w in T D(S y , x) and that (u, w) is not a directed edge on that path. Then k ≥ 1 and, so, v 1 ∈ {u, w}. It suffices to show that (v 1 , w) is a directed edge of T D (S y , x).
Observe first that, in view of Property (P6), (w, u) is not a directed edge in T D(S y , x) as otherwise T D(S y , x) would contain a directed cycle. Combined with the definition of S y it follows that either x|uw is a δ-triplet or we have a δ-fork on {x, u, w}. In either case, δ(u, x) = δ(w, x) holds. Since (u, v 1 ) is a directed edge in T D(S y , x), we also have that xv 1 |u is a δ-triplet. Hence, δ(v 1 , x) = δ(x, u) = δ(w, x) and so we cannot have a δ-fork on {x, w, v 1 }. Since, in view of Property (P4), we cannot have a δ-tricycle on If the first equality holds, then v 1 x|w is a δ-triplet and, so, (w, v 1 ) is a directed edge in T D(S y , x). Consequently, the directed path v 1 , . . . , v k , w concatenated with that edge forms a directed cycle in T D(S y , x), which is impossible in view of Property (P6) holding. Thus, δ(w, v 1 ) = δ(v 1 , x) must hold. Consequently, wx|v 1 is a δ-triplet and, so, (v 1 , w) is an edge in T D(S y , x), as required.
To establish the main result of this section (Theorem 8.3), we need to be able to distinguish between the sets defined in lines 8 and 9 of algorithm Build-Cycle when given a symbolic 3-dissimilarity δ on X and the restriction δ| Y of δ to a subset Y ⊆ X with |Y | ≥ 3. To this end, we augment for a symbolic 3-dissimilarity κ on X the definition of those sets by writing S i (κ) rather than S i , i = y, z.
Observe first that if δ is level-1 representable and Y ⊆ X such that |Y | ≥ 3, then the restriction δ| Y of δ to Y is clearly level-1 representable. Indeed, a level-1 representation N (δ| Y ) of δ| Y can be obtained from a level-1 representation N (δ) of δ using the following 2-step process. First, remove all leaves in X − Y and their respective incoming edges from N (δ) and then suppress all resulting degree two vertices. Next, apply algorithm Transform to the resulting network. This begs the question of when level-1 representations of symbolic 3-dissimilarities on subsets of X give rise to a level-1 representation of a symbolic 3-dissimilarity on X. To answer this question which is the purpose of Theorem 8.3 we require the next result. Proposition 8.2. Let δ be a symbolic 3-dissimilarity on X. Then the following statements hold.
Proof. (i) The proposition is straight-forward to show for Properties (P1) and (P2), since they involve three and four elements of X, respectively. Note that to see Property (Pi), 3 ≤ i ≤ 8, we may assume without loss of generality that Properties (Pj), 3 ≤ j ≤ i − 1, are satisfied by δ. For ease of readability, we put S y := S y (δ). If δ does not satisfy Property (P3) then there exists a connected component C of C(δ) and δ-tricycles τ, τ ∈ V (C) such that δ(L(τ )) = δ(L(τ )). Without loss of generality, we may assume that τ and τ are adjacent. Then |L(τ ) ∩ L(τ )| = 2. Let x, y, z ∈ X has a single directed edge or (b) there exists some v 1 ∈ S y such that both (u, v 1 ) and (v 1 , v) are directed edges of T D(S y , x).
If Case (a) holds, then δ| Z does not satisfy Property (P8) for Z = {x, y, z, u, v}. So assume that Case (b) holds. Then δ| Z does not satisfy Property (P8) for Z = {x, y, z, u, v, v 1 }. Since the definition of T D(S y , x) implies that xv|v 1 is a δ-triplet, it follows that δ(x, v) = δ(x, v 1 ). Hence, either δ(v, x) = δ(v, z) or δ(v 1 , x) = δ(v, z). By Properties (P3) and (P4) it follows in the first case that x||vz is a δ-tricycle, and that x||v 1 z is a δ-tricycle in the second case. Thus, either v or v 1 can play the role of y in τ . Consequently, δ restricted to Z = Z − {y} does not satisfy Property (P8).
(ii) This is a straight-forward consequence of Theorem 7.1 and Proposition 8.2(i). Proof. Suppose first that δ is level-1 representable. Then, by the observation preceding Proposition 8.2, δ| Y is level-1 representable, for all subsets Y ⊆ X of size |X| − 1.
Conversely, suppose that X is such that for all subsets Y ⊆ X of size |X| − 1, the restriction δ| Y is level-1 representable but that δ is not level-1 representable. Then, by Proposition 8.2 there exists a subset Y ⊆ X with |Y | ∈ {3, 4, 5} such that δ| Y is also not level-1 representable. But then δ restricted to any subset Z of X size |X| − 1 that contains Y also is not level-1 representable which is impossible.

Conclusion
In this paper, we have introduced the novel Network-Popping algorithm. It takes as input an orthology relation, formalized as a symbolic 3-dissimilarity δ, and finds, in polynomial time, a level-1 representation of δ precisely if such a representation exists. In addition to this representation being a discriminating symbolic representation of δ precisely if such a tree is supported by δ, Network-Popping enjoys several other attractive properties. As part of our analysis of Network-Popping, we characterize level-1 representable symbolic 3-dissimilarities δ in terms of eight natural properties that δ must satisfy. Last-but-not-least, we also characterize a level-1 representable symbolic 3-dissimilarity δ on some set X with |X| ≥ 6 in terms of level-1 representable orthology relations induced by δ on subsets of X of size |X| − 1. Combined with the polynomial run-time of Network-Popping this suggests that it could potentially be applied to large data sets within a Divide-and-Conquer framework thus providing an alternative to tree-based reconciliation or error correction approaches for orthology relations.
However a number of open questions remain. For example can other types of phylogenetic networks be used to also represent orthology relations. Interesting types of such networks might be tree-child networks [14] as they are uniquely determined by the trinets they induce and also regular networks [15] as they are known to be uniquely determined by the phylogenetic trees they induce, a property that is not shared by phylogenetic networks in general [4]. For those networks it would also be interesting to understand how the representation of an orthology relation in terms of those trees relates to the way such a relation is represented by the labelled network displaying the trees. Motivated by the point made in [3, Chapter 12] on k-estimates, k ≥ 3, already mentioned above it might also be interesting to investigate if symbolic k-dissimilarities for k ≥ 4 lend themselves as a useful formalization of orthology relations.
A further question concerns the fact that by evoking parsimony we only distinguish between 3 types of trinets associated to an orthology relation. Thus it might be interesting to investigate what can be done if this framework is replaced by e. g. a probabilistic one which assigns probability values to the trinets.