Abstract
Background
Supertree methods synthesize collections of small phylogenetic trees with incomplete taxon overlap into comprehensive trees, or supertrees, that include all taxa found in the input trees. Supertree methods based on the well established RobinsonFoulds (RF) distance have the potential to build supertrees that retain much information from the input trees. Specifically, the RF supertree problem seeks a binary supertree that minimizes the sum of the RF distances from the supertree to the input trees. Thus, an RF supertree is a supertree that is consistent with the largest number of clusters (or clades) from the input trees.
Results
We introduce efficient, local search based, hillclimbing heuristics for the intrinsically hard RF supertree problem on rooted trees. These heuristics use novel nontrivial algorithms for the SPR and TBR local search problems which improve on the time complexity of the best known (naïve) solutions by a factor of Θ(n) and Θ(n^{2}) respectively (where n is the number of taxa, or leaves, in the supertree). We use an implementation of our new algorithms to examine the performance of the RF supertree method and compare it to matrix representation with parsimony (MRP) and the triplet supertree method using four supertree data sets. Not only did our RF heuristic provide fast estimates of RF supertrees in all data sets, but the RF supertrees also retained more of the information from the input trees (based on the RF distance) than the other supertree methods.
Conclusions
Our heuristics for the RF supertree problem, based on our new local search algorithms, make it possible for the first time to estimate large supertrees by directly optimizing the RF distance from rooted input trees to the supertrees. This provides a new and fast method to build accurate supertrees. RF supertrees may also be useful for estimating majorityrule() supertrees, which are a generalization of majorityrule consensus trees.
Introduction
Supertree methods provide a formal approach for combining small phylogenetic trees with incomplete species overlap in order to build comprehensive species phylogenies, or supertrees, that contain all species found in the input trees. Supertree analyses have produced the first familylevel phylogeny of flowering plants [1] and the first phylogeny of nearly all extant mammal species [2]. They have also enabled phylogenetic analyses using largescale genomic data sets in bacteria, across eukaryotes, and within plants [3,4] and have helped elucidate the origin of eukaryotic genomes [5]. Furthermore, supertrees have been used to examine rates and patterns of species diversification [1,2], to test hypotheses regarding the structure of ecological communities [6], and to examine extinction risk in current species [7].
Although supertrees can support largescale evolutionary and ecological analyses, there are still numerous concerns about the performance of existing supertree methods (e.g., [814]). In general, an effective supertree method must accurately estimate phylogenies from large data sets in a reasonable amount of time while retaining much of the phylogenetic information from the input trees.
By far the most commonly used supertree method is matrix representation with parsimony (MRP), which works by solving the parsimony problem on a binary matrix representation of the input trees [15,16]. While the parsimony problem is NPhard, MRP can take advantage of fast and effective hillclimbing heuristics implemented in PAUP* or TNT (e.g., [1719]). MRP heuristics often perform well in analyses of both simulated and empirical data sets (e.g., [2022]); however, there are numerous criticisms of MRP. For example, MRP shows evidence of biases based on the shape and size of input trees [8,11], and MRP supertrees may contain relationships that are not supported by any of the input trees [9,12]. Furthermore, it is unclear if or why minimizing the parsimony score of a matrix representation of input trees is a good optimality criterion or should produce accurate supertrees.
Since evolutionary biologists rarely, if ever, know the true relationships for a group of species, it is difficult to assess the accuracy of supertree, or any phylogenetic, methods. One approach to evaluate the accuracy of supertrees is with simulations (e.g., [20,21]). However, simulations inherently simplify the true processes of evolution, and it is unclear how well the performance of a phylogenetic method in simulations corresponds to its performance with empirical data. Perhaps a more useful way to define the accuracy of a supertree method is to quantify the amount of phylogenetic information from the input trees that is retained in the supertree. Ideally, we want the supertree to reflect the input tree topologies as much as possible. This suggests that the supertree objective should directly evaluate the similarity of the supertree to the input trees (e.g., [11,23,24]).
Numerous metrics exist to measure the similarity of input trees to a supertree, and the RobinsonFoulds (RF) distance metric [25] is among the most widely used. In fact, numerous studies have evaluated the performance of supertree methods, including MRP, by measuring the RF distance between collections of input trees and the resulting supertrees (e.g., [11,20,21]). The RF supertree problem seeks a binary supertree that minimizes the sum of the RF distances between every rooted input tree and the supertree. The intuition behind seeking a binary supertree is that, in this setting, minimizing the RF distance is equivalent to maximizing the number of clusters (or clades) that are shared by the supertree and the input trees. Thus, an RF supertree is a supertree that is consistent with the largest number of clusters from the input trees. Unfortunately, as with MRP, computing RF supertrees is NPhard [26]. In this work, we describe efficient hillclimbing heuristics to estimate RF supertrees. These heuristics allow the first largescale estimates of RF supertrees and comparisons of the accuracy of RF supertrees to other commonly used supertree methods.
The RF distance metric between two rooted trees is defined to be a normalized count of the symmetric difference between the sets of clusters of the two trees. In the supertree setting, the input trees will often have only a strict subset of the taxa present in the supertree. Thus, a high RF distance between an input tree and a supertree does not necessarily correspond to conflicting evolutionary histories; it can also indicate incomplete phylogenetic information. Consequently, in order to compute the RF distance between an input tree which has only a strict subset of the taxa in the supertree, we first restrict the supertree to only the leaf set of the input tree. This adapted version of the RF distance is not a metric, or even a distance measure (mathematically speaking). However, for convenience, we will refer to this adapted version of the RF distance metric using the same name.
Previous work
Supertree methods are a generalization of consensus methods, in which all the input trees have the same leaf set. The problem of finding an optimal median tree under the RF distance in such a consensus setting is wellstudied. In particular, it is known that the majorityrule consensus of the input trees must be a median tree [27], and it can be found in polynomial time. On the other hand, finding the optimum binary median tree, i.e. an RF supertree, in the consensus setting is NPhard [26]. This implies that computing an RF supertree in general is NPhard as well.
Our definition of RF distance between two trees where one has only a strict subset of the taxa in the other, corresponds to the distance measure used to define "majorityrule() supertrees" by Cotton and Wilkinson [28]. This definition restricts the larger tree to only the leaf set of the smaller tree before evaluating the RF distance. Majorityrule() supertrees are defined to be the strict consensus of all the optimal median trees under the RF distance. These median trees are defined similarly to RF supertrees, except that RF supertrees must be binary while the median trees can be nonbinary. In general, majorityrule supertrees [28], in both their () and (+) variants, seek to generalize the majorityrule consensus. Indeed, majorityrule supertrees have been shown to have several desirable properties reminiscent of majorityrule consensus trees [29]. Although majorityrule supertrees and RF supertrees are both based on minimizing RF distance, they represent two different approaches to supertree construction. In particular, the RF supertree method seeks a supertree that is consistent with the largest number of clusters (clades) from the input trees, while majorityrule supertrees do not. Nevertheless, as we discuss later, RF supertrees could be used as a starting point to estimate majorityrule() supertrees.
The RF distance between two trees on the same size n leaf set, with leaves labeled by integers {1, ..., n}, can be computed in O(n) time [30]. In fact, an (1 + ϵ)approximate value of the RF distance can be computed in sublinear time, with high probability [31].
In the case of unrooted trees, the RF distance metric is sometimes also known as the splits metric (e.g., [32]). The supertree analysis package Clann [23] provides heuristics that operate on unrooted trees and attempt to maximize the number of splits shared between the input trees and the inferred supertree. This method is called the "maximum splitsfit" method.
Local Search
We use a heuristic approach for the RF supertree problem. Local search is the basis of effective heuristics for many phylogenetic problems. These heuristics iteratively search through the space of possible supertrees guided, at each step, by solutions to some local search problem. More formally, in these heuristics, a tree graph (see [32,33]) is defined for the given set of input trees and some fixed tree edit operation. The node set of this tree graph represents the set of all supertrees on the given input trees. An edge is drawn between two nodes exactly if the corresponding trees can be transformed into each other by one tree edit operation. In our setting, the cost of a node in the graph is the RF distance between the supertree represented by that node and the given input trees. Given an initial node in the tree graph, the heuristic's task is to find a maximallength path of steepest descent in the cost of its nodes and to return the last node on such a path. This path is found by solving the local search problem at every node along the path. The local search problem is to find a node with the minimum cost in the neighborhood of a given node. The neighborhood is defined by some tree edit operation, and hence, the time complexity of the local search problem depends on the tree edit operation used.
Two of the most extensively used tree edit operations for supertrees are rooted Subtree Prune and Regraft (SPR) [3335] and rooted Tree Bisection and Reconnection (TBR) [22,33,34]. The best known (naïve) algorithms for the SPR and TBR local search problems for the RF supertree problem require O(kn^{3}) and O(kn^{4}) time respectively, where k is the number of input trees, and n is the number of leaves in the supertree solution.
Our Contribution
We describe efficient hillclimbing heuristics for the RF supertree problem. These heuristics are based on novel nontrivial algorithms that can solve the corresponding local search problems for both SPR and TBR in O(kn^{2}) time, yielding speedups of Θ(n) and Θ(n^{2}) over the best known solutions respectively. These new algorithms are inspired by fast local search algorithms for the gene duplication problem [36,37]. Note that while the supertree itself must be binary, our algorithms work even if the input trees are not. We also examine the performance of the RF supertree method using four published supertree data sets, and compare its performance with MRP and the triplet supertree method [38]. We demonstrate that the new algorithms enable RF supertree analyses on large data sets and that the RF supertree method outperforms other supertree methods in finding supertrees that are most similar to the input trees based on the RF distance metric.
Basic Notation and Preliminaries
A tree T is a connected acyclic graph, consisting of a node set V (T) and an edge set E(T). T is rooted if it has exactly one distinguished node called the root which we denote by rt(T). Throughout this work, the term tree refers to a rooted tree. We define ≤_{T }to be the partial order on V (T) where x ≤_{T }y if y is a node on the path between rt(T) and x. The set of minima under ≤_{T }is denoted by ℒ(T) and its elements are called leaves. The set of all nonroot internal nodes of T, denoted by I(T), is defined to be the set V (T)\(ℒ(T) ∪ {rt(T)}). If {x, y} ∈ E(T) and x ≤_{T}y then we call y the parent of x denoted by pa_{T }(x) and we call x a child of y. The set of all children of y is denoted by Ch_{T}(y). T is fully binary if every node has either zero or two children. If two nodes in T have the same parent, they are called siblings. The least common ancestor of a nonempty subset L ⊆ V (T), denoted as lca(L), is the unique smallest upper bound of L under ≤_{T}. The subtree of T rooted at node y ∈ V(T), denoted by T_{y}, is the tree induced by {x ∈ V (T): x ≤ y}. For each node v ∈ I(T), the cluster (v) is defined to be the set of all leaf nodes in T_{v}; i.e. (v) = ℒ(T_{v}). We denote the set of all clusters of a tree T by ℋ(T). Given a set L ⊆ ℒ(T), let T' be the minimal rooted subtree of T with leaf set L. We define the leaf induced subtree T [L] of T on leaf set L to be the tree obtained from T' by successively removing each nonroot node of degree two and adjoining its two neighbors. The symmetric difference of two sets A and B, denoted by AΔB, is the set (A\B) ∪ (B\A). A profile is a tuple of trees (T_{1}, ..., T_{k}).
The RF Supertree Problem
Given a profile , we define a supertree on to be a fully binary tree T* where .
Definition 1 (RF Distance). Given a profile = (T_{1}, ..., T_{k}) and a supertree T * on , we define the RF distance as follows:
1. For any T_{i}, where 1 ≤ i ≤ k, RF (T_{i}, T*) = ℋ(T_{i})Δℋ(T*[ℒ(T_{i})]).
2.
3. Let be the set of supertrees on , then .
Remark: Traditionally, the value of the RF distance, as computed above, is normalized by multiplying by 1/2. However, this does not affect the definition or computation of RF supertrees, and therefore, we do not normalize the RF distance.
Problem 1 (RF Supertree).
Instance: A profile .
Find: A supertree T_{opt}on such that RF (, T_{opt}) = RF ().
Recall that the RF Supertree problem is NPhard [27].
Local Search Problems
Here we first provide definitions for the rerooting operation (denoted RR) and the TBR[22] and SPR[35] edit operations and then formulate the related local search problems that were motivated in the introduction.
Definition 2 (RR operation). Let T be a tree and x ∈ V (T). RR(T, x) is defined to be the tree T, if x = rt(T) or x ∈ Ch(rt(T)). Otherwise, RR(T, x) is the tree that is obtained from T by (i) suppressing rt(T), and (ii) subdividing the edge {pa(x), x} by a new root node. We define the following extension: RR(T) = ∪_{x ∈ V(T)}{RR(T, x)}.
For technical reasons, before we can define the TBR operation, we need the following definition.
Definition 3 (Planted tree). Given a tree T, the planted tree Φ(T) is the tree obtained by adding a root edge {p, rt(T)}, where p ∉ V (T), to T.
Definition 4 (TBR operation). (See Fig. 1) Let T be a tree, e = (u, v) ∈ E(T), where u = pa(v), and X, Y be the connected components that are obtained by removing edge e from T where v ∈ X and u ∈ Y. We define TBR_{T}(v, x, y) for x ∈ X and y ∈ Y to be the tree that is obtained from Φ(T) by first removing edge e, then replacing the component X by RR(X, x), and then adjoining a new edge f between x' = rt(RR(X, x)) and Y as follows:
Figure 1. TBR Operation. Example depicting a TBR operation which transforms tree S into tree S' = TBR_{S}(v, x, y).
1. Create a new node y' that subdivides the edge (pa(y), y).
2. Adjoin the edge f between nodes x' and y'.
3. Suppress the node u, and rename x' as v and y' as u.
4. Contract the root edge.
Notation. We define the following:
1. TBR_{T }(v, x) = ∪_{y ∈ Y }{TBR_{T }(v, x, y)}
2. TBR_{T }(v) = ∪_{x ∈ X }TBR_{T }(v, x)
3. TBR_{T }= ∪_{(u, v) ∈ E(T) }TBR_{T}(v)
Definition 5 (SPR operation). Let T be a tree, e = (u, v) ∈ E(T), where u = pa(v), and X, Y be the connected components that are obtained by removing edge e from T where v ∈ X and u ∈ Y. We define SPR_{T }(v, y), for y ∈ Y, to be the tree TBR_{T }(v, v, y). We say that the tree SPR_{T }(v, y) is obtained from T by a subtree prune and regraft (SPR) operation that prunes subtree T_{v }and regrafts it above node y.
Notation. We define the following:
1. SPR_{T }(v) = ∪_{y ∈ Y}{SPR_{T }(v, y)}
2. SPR_{T }= ∪_{(u, v) ∈ E(T) }SPR_{T}(v)
Note that an SPR operation for a given tree T can be briefly described through the following four steps: (i) prune some subtree P from T, (ii) add a root edge to the remaining tree S, (iii) regraft P into an edge of the remaining tree S, and (iv) contract the root edge.
We now define the relevant local search problems based on the TBR and SPR operations.
Problem 2 (TBRScoring (TBRS)). Given instance ⟨, T⟩, where is the profile (T_{1}, ..., T_{k}) and T is a supertree on , find a tree T* ∈ TBR_{T }such that .
Problem 3 (TBRRestricted Scoring (TBRRS)). Given instance ⟨, T, v⟩, where is the profile (T_{1}, ..., T_{k}), T is a supertree on , and v is a nonroot node in V (T), find a tree T * ∈ TBR_{T }(v) such that .
The problems SPRScoring (SPRS) and SPRRestricted Scoring (SPRRS) are defined analogously to the problems TBRS and TBRRS respectively.
Throughout the remainder of this manuscript, k is the number of trees in the profile , T denotes a supertree on , and n is the number of leaves in T. The following observation follows from Definition 4.
Observation 1. The TBRS problem on instance ⟨, T⟩ can be solved by solving the TBRRS problem E(T) times.
We show how to solve the TBRS problem on the instance ⟨, T⟩ in O(kn^{2}) time. Since SPR_{T }⊆ TBR_{T }this also implies an O(kn^{2}) solution for the SPRS problem. This gives a speedup of Θ(n^{2}) and Θ(n) over the best known (naïve) algorithms for the TBRS and SPRS problems respectively.
In particular, we first show that any instance of the TBRRS problem can be decomposed into an instance of an SPRRS problem, and an instance of a Rooting problem (defined in the next section). We show how to solve both these problems in O(kn) time, yielding an O(kn) time solution for the TBRRS problem. This immediately implies an O(kn^{2}) time algorithm for the TBRS problem (see Observation 1).
Note that the size of the set TBR_{T }is Θ(n^{3}). Thus, for each tree in the input profile the time complexity of computing and enumerating the RF distances of all trees in TBR_{T }is Ω(n^{3}). However, to solve the TBRS problem one only needs to find a tree with the minimum RF distance. This lets us solve the TBRS problem in time that is sublinear in the size of TBR_{T}. In fact, after the initial O(kn^{2}) preprocessing step, our algorithm can output the RF distance of any tree in TBR_{T }in O(1) time.
Structural Properties
Throughout this section, we limit our attention to one tree S from the profile . We show how to solve the TBRRS problem for the instance ⟨(S), T, v⟩ for some nonroot node v ∈ V (T) in O(n) time. Based on this solution, it is straightforward to solve the TBRRS problem on the instance ⟨, T, v⟩ within O(kn) time as well. For clarity, we will also assume that ℒ(S) = ℒ(T). In general, if ℒ(S) ⊂ ℒ(T) then we can simply set T to be T [ℒ(S)]. This takes O(n) time and, consequently, does not affect the time complexity of our algorithm.
Our algorithm makes use of the LCA mapping from S to T. This mapping is defined as follows.
Definition 6 (LCA Mapping). Given two trees T' and T such that ℒ(T') ⊆ ℒ(T), the LCA mapping ℳ_{T', T}: V(T') → V(T) is the mapping ℳ_{T', T }(u) = lca_{T }(ℒ()).
Notation. We define a boolean function f_{T}: I(S) → {0, 1} such that f_{T }(u) = 1 if there exists a node v ∈ I(T) such that (u) = (v), and f_{T }(u) = 0 otherwise. Thus, f_{T }(u) = 1 if and only if the cluster (u) exists in the tree T as well. Additionally, we define ℱ_{T }= {u ∈ I(S): f_{T }(u) = 0}; that is, ℱ_{T }is the set of all nodes u ∈ I(S) such that the cluster (u) does not exist in the tree T.
The following lemma associates the value RF(S, T) with the cardinality of the set ℱ_{T}.
Lemma 1. RF(S, T) = I(T)  I(S) + 2·ℱ_{T }.
Proof. Let denote the set {u ∈ I(S): f_{T }(u) = 1}. By the definition of RF (S, T), we must have RF(S, T) = I(T) + I(S)  2·. And hence, since  + ℱ_{T } = I(S), we get RF(S, T) = I(T)  I(S) + 2·ℱ_{T }. □
Lemma 2. For any u ∈ I(S), f_{T }(u) = 1 if and only if (u) = (ℳ_{S, T }(u)).
Proof. If (u) = (ℳ_{S, T}(u)) then we must have (u) = (ℳ_{S, T }(u)) and, consequently, f_{T }(u) = 1. In the other direction, if (u) ≠ (ℳ_{S, T }(u)), then we must have (u) ⊂ (ℳ_{S, T }(u)) and, consequently, f_{T }(u) = 0. □
The LCA mapping from S to T can be computed in O(n) time [39], and consequently, by Lemmas 1 and 2, we can compute the RF distance between S and T in O(n) time as well (other O(n)time algorithms for calculating the RF distance are presented in [30,31]). Moreover, Lemma 1 implies that in order to find a tree T* ∈ TBR_{T }(v) such that , it is sufficient to find a tree T* ∈ TBR_{T }(v) for which .
Remark: An implicit assumption here is that the leaves of both trees are labeled by integers {1, ..., n}. If the leaf labels are arbitrary, then we require an additional O(kn log n)time preprocessing step to relabel the leaves of the trees in the given profile. Note, however, that this additional step does not add to the overall time complexity of solving the TBRS or SPRS problems.
We now show that the TBRRS problem can be solved by solving two smaller problems separately and combining their solutions.
As before, we limit our attention to one tree S from the profile . Given the TBRRS instance ⟨(S), T, v⟩, we define a bipartition {X, } of I(S), where X = {u ∈ I(S): ℳ_{S, T}(u) ∈ V (T_{v})}.
Lemma 3. If u ∈ X, then f_{T'}(u) = f_{T }(u) for all T ' ∈ TBR_{T }(v, v). If u ∈ and y denotes the sibling of v, then f_{T'}(u) = f_{T }(u), where T' = TBR_{T}(v, x, y) for any x ∈ V (T_{v}).
Proof. Consider the case when u ∈ X. Let T' be any tree in TBR_{T}(v, v) and let node y ∈ V (T) be such that T' = TBR(v, v, y). Thus, for any node w ∈ V (T_{v}), the subtrees T_{v }and must be identical. Since u ∈ X, we must have ℳ_{S, T }(u) ∈ T_{v }and, consequently, . Lemma 2 now implies that f_{T' }(u) = f_{T }(u).
Now consider the case when u ∈ . Node y denotes the sibling of v in tree T and let T' = TBR(v, x, y), for some x ∈ V (T_{v}). Thus, for any node w ∈ V(T)\V(T_{v}), we must have ℒ_{T}(w) = ℒ_{T' }(w). Moreover, the leaf sets of the two subtrees rooted at the children of w in T must be identical to the leaf sets of the two subtrees rooted at the children of w in T': This implies that if ℳ_{S, T}(u) = w, then ℳ_{S, T' }(u) = w as well. By Lemma 2 we must therefore have f_{T' }(u) = f_{T }(u). □
Lemma 3 implies that a tree in TBR_{T }(v) with smallest RF distance can be obtained by optimizing the rooting for the pruned subtree, and optimizing the regraft location separately. This allows us to obtain a tree in TBR_{T }(v) with smallest RF distance by evaluating only O(n) trees. Contrast this with the naïve approach to finding a tree in TBR_{T }(v) with smallest total distance, which is to evaluate all trees obtained by rerooting the pruned subtree in all possible ways, and, for each rerooting, regrafting the subtree in all possible locations. Since there are O(n) ways to reroot the pruned subtree, and O(n) ways to regraft, this would require evaluating O(n^{2}) trees. It is interesting to note that this ability to decompose the TBRRS problem into two simpler problems is not unique to the context of RF supertrees alone. For example, it has been observed that a similar decomposition can be achieved in the context of the gene duplication problem [37].
Thus, to solve the TBRRS problem, we must find (i) a rerooting T' of the subtree T_{v }for which ℱ_{T' }is minimized, and (ii) a regraft location y for T_{v }which minimizes ℱ_{SPR}_{(v, y)}. Observe that the problem in part (ii) is simply the SPRRS problem on the input instance ⟨(S), T, v⟩. For part (i), consider the following problem statement.
Problem 4 (Rooting). Given instance ⟨, T, v⟩, where is the profile (T_{1}, ..., T_{k}), T is a supertree on , and v is a nonroot node in V (T), find a node x ∈ V (T_{v}) for which RF (, TBR_{T }(v, x, y)) is minimum, where y denotes the sibling of v in T.
Note that the problem in part (i) is the Rooting problem on the input instance ⟨(S), T, v⟩. We show how to solve both the Rooting and the SPRRS problems in O(n) time on instance ⟨(S), T, v⟩. As seen above, based on Lemma 3, this immediately implies that the TBRRS problem for a profile consisting of a single tree can be solved in O(n) time. To solve the TBRRS problem on instance ⟨, T, v⟩, we simply solve the Rooting and SPRRS problems separately on the input instance ⟨, T, v⟩, which takes O(kn) time (see Theorems 3 and 4). We thus have the following two theorems.
Theorem 1. The TBRRS problem can be solved in O(kn) time.
Theorem 2. The TBRS problem can be solved in O(kn^{2}) time.
Solving the Rooting Problem
To solve the Rooting problem on instance ⟨(S), T, v⟩, we rely on an efficient algorithm for computing the value of f_{T' }(u) for any T' ∈ RR(T_{v}) and any u ∈ I(S). This algorithm relies on the following five lemmas. Let a denote the node ℳ_{S, T }(u), y denote the sibling of v in T, and T' = TBR_{T}(v, x, y) for x ∈ V (T_{v}). Depending on a and f_{T }(u) there are five possible cases: (i) a ∉ V (T_{v}), (ii) a = rt(T_{v}) and f_{T }(u) = 1, (iii) a = rt(T_{v}) and f_{T }(u) = 0, (iv) a ∈ V (T_{v})\rt(T_{v}) and f_{T }(u) = 1, and (v) a ∈ V (T_{v})\rt(T_{v}) and f_{T }(u) = 0. Lemmas 4 through 8 characterize the value f_{T' }(u) for each of these five cases respectively.
Lemma 4. If a ∉ V (T_{v}), then f_{T' }(u) = f_{T }(u) for any x ∈ V (T_{v}).
Proof. Follows directly from Lemma 3. □
Lemma 5. If a = rt(T_{v}) and f_{T }(u) = 1, then f_{T' }(u) = 1 for all x ∈ V (T_{v}).
Proof. Since we have a = rt(T_{v}) and f_{T }(u) = 1, by Lemma 2 we must have ℒ(S_{u}) = ℒ(T_{v}). Thus, for any x ∈ V (T_{v}), ℳ_{S, T' }(u) must be the root of the subtree RR(T_{v}, x). The lemma follows. □
Lemma 6. Let L denote the set ℒ(T_{v})\ℒ(S_{u}), and let (L). If a = rt(T_{v}) and f_{T }(u) = 0, then,
1. for b, f_{T' }(u) = 0, and,
2. for b, f_{T' }(u) = 1 if and only if L = ℒ(T_{b}).
Proof. Since a = rt(T_{v}) and f_{T }(u) = 0, by Lemma 2 we must have ℒ(S_{u}) ≠ ℒ(T_{v}). We analyze each part of the lemma separately.
1. b: For this case to be valid, we must have rt(T_{v}). Therefore, let (b). For any T' in this case, b' = pa_{T' }(b). Moreover, ℒ() ∩ ℒ(S_{u}) ≠ ∅. Therefore, we must have b <_{T' }ℳ_{S, T' }(u) Hence, in this case, and, consequently, Lemma 2 implies that f_{T' }(u) = 0.
2. b: We divide our analysis into two cases:
(a) L = ℒ(T_{b}): In this case we must have b ≠ rt(T_{v}). Therefore, let b' denote the parent of b in tree T_{v}. Now consider the tree T'. The set ℒ() must be identical to ℒ(S_{u}). Hence, f_{T' }(u) = 1 in this case.
(b) L ≠ ℒ(T_{b}): We claim that there does not exist any edge (pa(w), w) ∈ E(T_{v}) such that ℒ(T_{w}) is either ℒ(S_{u}) or L. Let us suppose, for the sake of contradiction, that such an edge exists. If ℒ(T_{w}) = ℒ(S_{u}) then we must have a = w, which is a contradiction since a = rt(T_{v}). If ℒ(T_{w}) = g L then we must have b = w, and, consequently, L ≠ ℒ(T_{b}), which is, again, a contradiction. Thus, such an edge (pa(w), w) ∈ E(T_{v}) cannot exist. Hence, we must have f_{T' }(u) = 0 for every x ∈ V (T_{v}) in this case.
The lemma follows. □
Lemma 7. If a ∈ V (T_{v})\rt(T_{v}) and f_{T}(u) = 1, then f_{T' }(u) = 0 if and only if x < T_{v }a.
Proof. By Lemma 2 we must have ℒ(S_{u}) = ℒ(T_{a}). We have two cases:
1. a: In this case we must have ℳ_{S, T' }(u) = a, and ℒ(T_{a}) = ℒ(). Thus, ℒ(S_{u}) = ℒ() and hence, f_{T' }(u) = 1.
2. a: In this case, ℳ_{S, T' }(u) must be the root of the subtree RR(T_{v}, x). Since ℒ(RR(T_{v}, x)) = ℒ(T_{v}), and ℒ(S_{u}) ≠ ℒ(T_{v}), Lemma 2 implies that f_{T' }(u) = 0.
The lemma follows. □
Lemma 8. If a ∈ V (T_{v})\rt(T_{v}) and f_{T }(u) = 0, then f_{T' }(u) = 0 for all x ∈ V (T_{v}).
Proof. By Lemma 2 we must have ℒ(S_{u}) ≠ ℒ(T_{a}). We have two possible cases:
1. a: In this case we must have ℳ_{S, T' }(u) = a, and ℒ(T_{a}) = ℒ(). Thus, ℒ(S_{u}) ≠ ℒ() and hence, f_{T' }(u) = 0
2. a: In this case, ℳ_{S, T' }(u) must be the root of the subtree RR(T_{v}, x). Since ℒ(RR(T_{v}, x)) = ℒ(T_{v}), and ℒ(S_{u}) ≠ ℒ(T_{v}), Lemma 2 implies that f_{T' }(u) = 0.
The lemma follows.
The Algorithm. For any x ∈ V (T_{v}) let A(x) denote the cardinality of the set
{u ∈ I(S): f_{T }(u) = 0, but f_{T' }(u) = 1}, and B(x) the cardinality of the set
{u ∈ I(S): f_{T }(u) = 1, but f_{T' }(u) = 0}, where T' = TBR_{T }(v, x, y).
By definition, to solve the Rooting problem we must find a node x ∈ V (T_{v}) for which A(x)  B(x) is maximized. Our algorithm computes, at each node x ∈ V (T_{v}), the values A(x) and B(x).
In a preprocessing step, our algorithm computes the mapping ℳ_{S, T }as well as the size of each cluster in S and T, and creates and initializes (to 0) two counters α(x) and β(x) at each node x ∈ V (T_{v}). This takes O(n) time. When the algorithm terminates, the values α(x) and β(x) at any x ∈ V (T_{v}) will be the values α(x) and β(x).
Recall that, given u ∈ I(S), a denotes the node ℳ_{S, T }(u). Thus, any given u ∈ I(S) must satisfy the precondition (given in terms of a) of exactly one of the the Lemmas 4 through 8. Moreover, the precondition of each of these lemmas can be checked in O(1) time.
The algorithm then traverses through S and considers each node u ∈ I(S). There are three cases:
1. If u satisfies the preconditions of Lemmas 4, 5, or 8 then we must have f_{T' }(u) = f_{T }(u). Consequently, we do nothing in this case.
2. If u satisfies the precondition of Lemma 7, then we increment the value of β(x) at each node x ∈ V (T_{a})\{a} (where a is as in the statement of Lemma 7). To do this efficiently we can simply increment a counter at node a such that, after all u ∈ I(S) have been considered, a single preorder traversal of T_{v }can be used to compute the correct values of β(x) at each x ∈ V (T_{v}).
3. If u satisfies the precondition of Lemma 6, then we proceed as follows: Let a and L be as in the statement of Lemma 6. According to the Lemma, if we can find a node b ∈ V (T_{v}) such that (L) and ℒ(T_{b}) = L, then we increment the value of α(x) at each node x ∈ V (T_{b}); otherwise, if such a b does not exist, we do nothing. As before, to do this efficiently, we only increment a single counter at node b such that, after all u ∈ I(S) have been considered, a preorder traversal of T_{v }suffices to compute the correct values of α(x) at each x ∈ V (T_{v}). In order to prove the O(n) runtime for this algorithm we will now explain how to precompute such a corresponding node b (if it exists), for each u ∈ I(S) satisfying the precondition of Lemma 6, within O(n) time. Note that any edge in a tree bipartitions its leaf set. Construct the tree S' = S [ℒ(T_{v})]. Observe that, given any candidate u, the corresponding node b exists if and only if the partition of ℒ(S') induced by the edge (u, pa(u)) E(S'), is also induced by some edge, e, in the tree T_{v }If such an e exists, then b must be that node on e which is farther away from the root, i.e. the edge e must be the edge (b, pa(b)) in T_{v }This edge e (or its absence) can be precomputed, for all candidate u, as follows: Compute the strict consensus of the unrooted variants of the trees S' and T_{v}. Every edge in this strict consensus corresponds to an edge in S' and an edge in T_{v }that induce the same bipartitions in the two trees.
Thus, for all candidate u that lie on such an edge, the corresponding node b can be inferred in O(1) time (by using the association between the edges of the strict consensus and the edges of S' and T_{v}), and for all candidate u that do not lie on such an edge, we know that the corresponding node b does not exist. This strict consensus of the unrooted variants of S' and T_{v }can be precomputed within O(n) time by using the algorithm of Day [30].
Hence, the Rooting problem for a profile consisting of a single tree can be solved in O(n) time; yielding the following theorem.
Theorem 3. The Rooting problem can be solved in O(kn) time.
Solving the SPRRS Problem
We will show how to solve the SPRRS problem on instance ⟨(S), T, v⟩ in O(n) time. Consider the tree R = SPR_{T }(v, rt(T)) Observe that, since SPR_{R}(v) = SPR_{S}(v), solving the SPRRS problem on instance ⟨(S), T, v⟩ is equivalent to solving it on the instance ⟨(S), R, v⟩. Thus, in the remainder of this section, we will work with tree R instead of tree S. The following four lemmas let us efficiently infer, for any u ∈ I(S), whether f_{T' }(u) = 1 or f_{T' }(u) = 0, for any given T' ∈ SPR_{R}(v).
For brevity, let a denote the node ℳ_{S, R}(u), and let Q denote the set V (R)\(V (R_{v}) ∪ {rt(R)}). Let T' = SPR_{R}(v, x), for any x ∈ Q.
Depending on a and f_{R}(u) there are four possible cases: (i) a ∈ V (R_{v}), (ii) a ∈ Q and f_{R}(u) = 1, (iii) a ∈ Q and f_{R}(u) = 0, and (iv) a = rt(R). Lemmas 9 through 12 characterize the value f_{T' }(u) for each of these four cases respectively. □
Lemma 9. If a ∈ V (R_{v}), then f_{T' }(u) = f_{R}(u) for any x ∈ Q.
Proof. Observe that TBR_{R}(v, v) = SPR_{R}(v). Lemma 3 now immediately completes the proof. □
Lemma 10. If a ∈ Q and f_{R}(u) = 1, then,
1. f_{T' }(u) = 0, for x <_{R }a, and
2. f_{T' }(u) = 1, otherwise.
Proof. Since f_{R}(u) = 1, Lemma 2 implies that (u) = (a). Let T' = SPR_{R}(v, x); we now have two cases.
1. x <_{R }a: In this case ℳ_{S, T' }(u) = a, and, since (u) < (a) < (a), we must have f_{T' }(u) = 0 (by Lemma 2).
2. x ≮_{R }a: In this case ℳ_{S, T' }(u) = a, and since (u) = (a) = (a), we must have f_{T' }(u) = 1 (by Lemma 2).
The lemma follows. □
Lemma 11. If a ∈ Q and f_{R}(u) = 0, then f_{T' }(u) = 0 for any x ∈ Q.
Proof. Since f_{R}(u) = 0, Lemma 2 implies that (u) ≠ (a). Thus, by the definition of LCA mapping, (u) < (a). Let T' = SPR_{R}(v, x); we now have two cases.
1. x <_{R }a: In this case ℳ_{S, T' }(u) = a, and, since (u) < (a) < (a), we must have f_{T' }(u) = 0 (by Lemma 2).
2. x ≮_{R }a: In this case ℳ_{S, T' }(u) = a, and since (u) < (a) = (a), we must have f_{T' }(u) = 0 (by Lemma 2).
The lemma follows. □
For the next lemma, let S' be the tree obtained from S by suppressing all nodes s for which ℳ_{S, R}(s) ∈ R_{v}.
Lemma 12. If a = rt(R) and b = ℳ_{S', R}(u), then, f_{T' }(u) = 1 if and only if x <_{R }b and ℒ(R_{b}) + ℒ(R_{v}) = ℒ(S_{u}).
Proof. First, observe that, since a = rt(R), the mapping ℳ_{S', R}(u) is well defined. Second, since b = ℳ_{S', R}(u), we must have ℒ() ⊆ ℒ(R_{b}), which implies that ℒ(S_{u}) ⊆ ℒ(R_{v}) ⊆ ℒ(R_{b}). We now have the following three cases:
1. x ≮_{R }b: In this case we must have ℳ_{S, T' }(u) = lca_{T' }(x, b). By Lemma 2 we know that f_{T' }(u) = 1 only if (u) = (ℳ_{S, T' }(u)). However, since we have ℒ(S_{u}) ⊆ ℒ(R_{v}) ⊆ ℒ(R_{b}), and x ≮_{R }b, we must have (u) < (ℳ_{S, T' }(u)); and hence, f_{T' }(u) = 0.
2. x <_{R }b and ℒ(R_{b}) + ℒ(R_{v}) ≠ ℒ(S_{u}): In this case we must have ℳ_{S, T' }(u) = b. Since ℒ(R_{b}) + ℒ(R_{v}) ≠ ℒ(S_{u}), we must have ℒ(S_{u}) ⊂ ℒ(R_{v}) ∪ ℒ(R_{b}), which implies that (u) < (ℳ_{S, T' }(u)). Thus, by Lemma 2, we must have f_{T' }(u) = 0.
3. x <_{R }b and ℒ(R_{b}) + ℒ(R_{v}) = ℒ(S_{u}): In this case we must have ℳ_{S, T' }(u) = b. Moreover, since ℒ(R_{b}) + ℒ(R_{v}) = ℒ(S_{u}), we must have (u) = (ℳ_{S, T' }(u)). Thus, by Lemma 2, we must have f_{T' }(u) = 1.
The lemma follows. □
The Algorithm. Note that SPR_{T }(v) = SPR_{R}(v) = ∪_{x ∈ Q }SPR_{R}(v, x). For any x ∈ Q, let A(x) = {u ∈ I(S): f_{R}(u) = 0, but f_{T' }(u) = 1}, and B(x) = {u ∈ I(S): f_{R}(u) = 1, but f_{T' }(u) = 0}, where T' = SPR_{R}(v, x). By definition, to solve the SPRRS problem on instance ⟨(S), T, v⟩ we must find a node x ∈ Q for which A(x)  B(x) is maximized. Our algorithm computes, at each node x ∈ Q, the values A(x) and B(x).
In a preprocessing step, our algorithm first constructs the tree R computes the mapping ℳ_{S, R }as well as the size of each cluster in S and R, and creates and initializes (to 0) two counters α(x) and β(x) at each node x ∈ Q. This takes a total of O(n) time. When the algorithm terminates, the values α(x) and β(x), at any x ∈ Q will be the values A(x) and B(x).
Recall that, given u ∈ I(S), a denotes the node ℳ_{S, R}(u). Thus, any given u ∈ I(S) must satisfy the precondition (given in terms of a) of exactly one of the the Lemmas 9 through 12. Moreover, the precondition of each of these lemmas can be checked in O(1) time.
The algorithm then traverses through S and considers each node u ∈ I(S). There are three cases:
1. If u satisfies the preconditions of Lemmas 9 or 11 then we must have f_{T' }(u) = f_{R}(u) Consequently, we do nothing in this case.
2. If u satisfies the precondition of Lemma 10, then we increment the value of β(x) at each node x ∈ V (T_{a})\{a} (where a is as in the statement of Lemma 10). This can be done efficiently as shown in part (2) of the algorithm for the Rooting problem.
3. If u satisfies the precondition of Lemma 6, and if ℒ(R_{b}) + ℒ(R_{v}) = ℒ(S_{u}), then we increment the value of α(x) at each node x ∈ V (T_{b})\{b} (where a and b are as in the statement of Lemma 6).
Again, to do this efficiently, we increment a counter at node b, and perform a subsequent preorder traversal. Note also that the mapping ℳ_{S', R }can be computed in O(n) time in the preprocessing step, and hence the node b can be inferred in O(1) time. The condition ℒ(R_{b}) + ℒ(R_{v}) = ℒ(S_{u}) is also verifiable in O(1) time.
Hence, the SPRRS problem for a profile consisting of a single tree can be solved in O(n) time; yielding the following theorem.
Theorem 4. The SPRRS problem can be solved in O(kn) time.
Remark. To improve the performance of local search heuristics in phylogeny construction, the starting tree for the first local search step is often constructed using a greedy 'stepwise addition' procedure. This greedy procedure builds a starting species tree stepbystep by adding one taxon at a time at its locally optimal position. In the context of RF supertrees, our algorithm for the SPRRS problem also yields a Θ(n) speedup over naïve algorithms for this greedy procedure.
Experimental Evaluation
In order to evaluate the performance of the RF supertree method, we implemented an RF heuristic based on the SPR local search algorithm. We focused on the SPR local search because it is faster and simpler to implement than TBR, and in analyses of MRF and triplet supertrees, the performance of SPR and TBR was very similar [22,38]. We compared the performance of the RF supertree heuristic to MRP and the triplet supertree method (which seeks a supertree with the most shared triplets with the collection of input trees) using published supertree data sets from sea birds [40], marsupials [41], placental mammals [42], and legumes [43]. The published data sets contain between 7 and 726 input trees and between 112 and 571 total taxa (Table 1).
Table 1. Experimental Results.
There are a number of ways to implement any local search algorithm. Preliminary analyses of the RF heuristic based on the SPR local search indicated that, as with other phylogenetic methods, the starting tree can affect the estimate of the final supertree. Occasionally the SPR searches got caught in local optima with relatively high RFdistance scores. To ameliorate this potential problem, we implemented a ratchet search heuristic for RF supertrees based on the parsimony ratchet [44]. In general, a ratchet search performs a number of iterations  in our case 25  that consist of two local SPR searches: one in which the characters (input trees) are equally weighted, and another in which the set of the characters are reweighted. We reweighted the characters by randomly removing approximately twothirds of the input trees. The goal of reweighting the characters is to alter the tree space to avoid getting caught in a globally suboptimal part of the tree space. At the end of each iteration, the best tree is taken as the starting point of the next iteration. For each data set, we started RF ratchet searches from 20 random sequence addition starting trees, and we also ran three replicates starting from an optimal MRP supertree. All RF supertree analyses were performed on an 3 GHz Intel Pentium 4 based desktop computer with 1 GB of main memory. The RFratchet runs took between 5 seconds (for the Sea Birds data set) and 90 minutes (for the legume dataset) when starting from a random sequence addition tree. RFratchet runs starting from optimal MRP trees were at least twice as fast because they required fewer search steps.
For our MRP analyses, we also tried two heuristic search methods, both implemented using PAUP* [18]. First, we performed 20 replicates of TBR branch swapping from trees built with random addition sequence starting trees. Next, we performed 20 replicates of a parsimony ratchet search with TBR branch swapping. Based on the results of trial analyses, each ratchet search consisted of 25 iterations, each reweighting 15% of the characters. The PAUP* command block for the parsimony ratchet searches was generated using PAUPRat [45]. For each data set, we performed 20 replicates of a TBR local search heuristic starting with random addition sequence trees. Triplet supertrees were constructed using the program from Lin et al. [38]. We were unable to perform ratchet searches with the existing triplet supertree software, and also, due to memory limitations, we were unable to perform triplet supertree analyses on the legume data set.
Our analyses demonstrate the effectiveness of our local search heuristics for the RF supertree problem. In all four data sets, RFratchet searches found the supertrees with the lowest total RF distance to the input trees (Table 1). MRP also generally performs well, finding supertrees with RF distances between 0.14% (placental mammals) and 3.3% (sea birds) higher than the best score found by the RF supertree heuristics (Table 1). The triplet supertree method performs as well as the RF supertree method on the small sea bird data set; however, the triplet supertrees for the marsupial and placental mammal data sets have a much higher RF distance to the input trees than either the RF or MRP supertrees (Table 1). For all the data sets, the MRP supertrees had the lowest (best) parsimony score based on a binary matrix representation of the input trees (Table 1). Thus, not surprisingly, it appears that optimizing based on the parsimony score or the triplet distance to the input trees does not optimize the similarity of the supertrees to the input trees based on the RF distance metric (see also [11,13]).
All of the data sets used in this analysis are from published studies that used MRP. Therefore, it is not surprising that MRP performed well (but see [46]). Still, our results demonstrate that MRP leaves some room for improvement. If the goal is to find the supertrees that are most similar to the collection of input trees, the RF searches ultimately provide better estimates than MRP (Table 1).
Interestingly, while the MRP trees tend to have relatively low RFdistance scores, in some cases, such as the legume data set, trees with low RFdistance scores have high parsimony scores (Table 1). Thus, parsimony scores are not necessarily indicative of RF score, and MRP and RF supertree optimality criteria are certainly not equivalent. Still, MRP trees appear to be useful as starting points for RF supertree heuristics. Indeed, in three of the four data sets, the best RF trees were found in ratchet searches beginning from MRP trees (Table 1).
Our program for computing RF supertrees is freely available (for Windows, Linux, and Mac OS X) at http://genome.cs.iastate.edu/CBL/RFsupertrees webcite
Discussion and Conclusion
There is a growing interest in using supertrees for largescale evolutionary and ecological analyses. Yet there are many concerns about the performance of existing supertree methods, and the great majority of published supertree analyses have relied on only MRP [47]. Since the goal of a supertree analysis is to synthesize the phylogenetic data from a collection of input trees, it makes sense that an effective supertree method should directly seek the supertree that is most similar to the input trees. Our new algorithms make it possible, for the first time, to estimate large supertrees by directly optimizing the RF distance from the supertree to the input trees.
There are numerous alternate metrics to compare phylogenetic trees besides the RF distance, and any of these can be used for supertree methods (see, for example, [11]). Triplet distance supertrees [11,48], quartetfit and quartet joining supertrees [11,24], maximum splitsfit supertrees [11], and most similar supertrees [49] are all, like RF supertrees, estimated by comparing input trees to the supertree using tree distance measures. All of these methods may provide different, and perhaps equally valid, perspectives on supertree accuracy. Based on our experimental analyses using the RF and triplet supertree method, optimizing the supertree based on different distance measures can result in very different supertrees (Table 1). In the future, it will be important to characterize and compare the performance of these methods in more detail (see, for example, [11,50]).
The results also suggest several future directions for research. Although heuristics guided by local search problems, especially SPR and TBR, have been very effective for many intrinsically difficult phylogenetic inference problems, our experiments indicate that the tree space for RF supertrees is complex. The ratchet approach and also starting from MRP trees appears to improve the performance in the four examples we tested (Table 1). However, more work is needed to identify the most efficient ways to implement our fast local search heuristics. Also, the use of alternative supertrees methods (other than MRP) to generate starting trees might result in a better global strategy to compute RF supertrees and this should be investigated further. We note that the ideas presented in [51] can be directly used to perform efficient NNIbased local searches for the RF supertree problem. In particular, we can show that heuristic searches for the RF supertree problem, which perform a total of p local search steps based on 1, 2, or 3NNI neighborhoods (see [51]), can all be executed in O(kn(n + p)) time; yielding speedups of Θ(min{n, p}), Θ(n·min{n, p}) and Θ(n^{2}·min{n, p}) for heuristic searches that are based on naïve algorithms for 1, 2 and 3NNI local searches respectively. It would also be interesting to see if heuristics based on TBR perform significantly better than those based on SPR in inferring RF supertrees.
In some cases it might be desirable to remove the restriction that the supertree be binary. In the consensus setting, such a median tree can be obtained within polynomial time [27]; however, finding a median RF tree in the supertree setting is NPhard [52]. One simple way to estimate a nonbinary median tree could be to first compute an RF supertree and then to iteratively (and perhaps greedily) contract those edges in the supertree that result in a reduction in the total RF distance. Thus, our algorithms may even be useful for roughly estimating majorityrule() supertrees [28], which are essentially the strict consensus of all optimal, not necessarily binary, median RF trees, and have several desirable properties [29]. These majorityrule() supertrees are also the strict consensus of all maximumlikelihood supertrees [53]. Also, there are several alternate forms of the RF distance metric that could be incorporated into our local search algorithms. For example, in order to account for biases associated with the different sizes of input trees, we could normalize the RF distance for each input tree, dividing the observed RF distance by the maximum possible RF distance based on the tree size. Similarly, we could incorporate either branch length data or phylogenetic support scores (bootstrap values or posterior probabilities) from the input trees into the RF distance in order to give more weight to partitions that are strongly supported or separated by long branches (e.g., [25,54]). Our current implementation essentially treats all branch lengths as one and all partitions as equal. The addition of branch length or support data may further improve the accuracy of the RF supertree method.
Competing interests
The authors declare that they have no competing interests.
Authors' contributions
MSB was responsible for algorithm design and program implementation, contributed to the experimental evaluation, and wrote major parts of the manuscript. JGB performed the experimental evaluation and the analysis of the results, and contributed to the writing of the manuscript. OE and DFB supervised the project and contributed to the writing of the manuscript. All authors read and approved the final manuscript.
Acknowledgements
We thank Harris Lin for providing software for the triplet supertree analyses. This work was supported in part by NESCent and by NSF grants DEB0334832 and DEB0829674. MSB was supported in part by a postdoctoral fellowship from the Edmond J. Safra Bioinformatics program at TelAviv university.
References

Davies TJ, Barraclough TG, Chase MW, Soltis PS, Soltis DE, Savolainen V: Darwin's abominable mystery: Insights from a supertree of the angiosperms.
Proceedings of the National Academy of Sciences of the United States of America 2004, 101(7):19041909. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

BinindaEmonds ORP, Cardillo M, Jones KE, Macphee RDE, Beck RMD, Grenyer R, Price SA, Vos RA, Gittleman JL, Purvis A: The delayed rise of presentday mammals.
Nature 2007, 446(7135):507512. PubMed Abstract  Publisher Full Text

Daubin V, Gouy M, Perriere G: A Phylogenomic Approach to Bacterial Phylogeny: Evidence of a Core of Genes Sharing a Common History.
Genome Res 2002, 12(7):10801090. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Burleigh JG, Driskell AC, Sanderson MJ: Supertree bootstrapping methods for assessing phylogenetic variation among genes in genomescale data sets.
Systematic Biology 2006, 55:426440. PubMed Abstract  Publisher Full Text

Pisani D, Cotton JA, McInerney JO: Supertrees disentangle the chimerical origin of eukaryotic genomes.
Mol Biol Evol 2007, 24(8):17521760. PubMed Abstract  Publisher Full Text

Webb CO, Ackerly DD, McPeek M, Donoghue MJ: Phylogenies and community ecology.
Ann Rev Ecol Syst 2002, 33:475505. Publisher Full Text

Davies TJ, Fritz SA, Grenyer R, Orme CDL, Bielby J, BinindaEmonds ORP, Cardillo M, Jones KE, Gittleman JL, Mace GM, Purvis A: Phylogenetic trees and the future of mammalian biodiversity.
Proceedings of the National Academy of Sciences 2008, 105(Supplement 1):1155611563. Publisher Full Text

Purvis A: A modification to Baum and Ragan's method for combining phylogenetic trees.

Pisani D, Wilkinson M: Matrix Representation with Parsimony, Taxonomic Congruence, and Total Evidence.
Systematic Biology 2002, 51:151155. PubMed Abstract  Publisher Full Text

BinindaEmonds ORP, Gittleman JL, Steel MA: The (super) tree of life: procedures, problems, and prospects.
Annual Review of Ecology and Systematics 2002, 33:265289. Publisher Full Text

Wilkinson M, Cotton JA, Creevey C, Eulenstein O, Harris SR, Lapointe FJ, Levasseur C, McInerney JO, Pisani D, Thorley JL: The shape of supertrees to come: Tree shape related properties of fourteen supertree methods.
Syst Biol 2005, 54:419432. PubMed Abstract  Publisher Full Text

Goloboff PA: Minority rule supertrees? MRP, Compatibility, and Minimum Flip may display the least frequent groups.
Cladistics 2005, 21(3):282294. Publisher Full Text

Wilkinson M, Cotton JA, Lapointe FJ, Pisani D: Properties of Supertree Methods in the Consensus Setting.
Syst Biol 2007, 56(2):330337. PubMed Abstract  Publisher Full Text

Day WH, McMorris F, Wilkinson M: Explosions and hot spots in supertree methods.
Journal of Theoretical Biology 2008, 253(2):345348. PubMed Abstract  Publisher Full Text

Baum BR: Combining Trees as a Way of Combining Data Sets for Phylogenetic Inference, and the Desirability of Combining Gene Trees.
Taxon 1992, 41:310. Publisher Full Text

Ragan MA: Phylogenetic inference based on matrix representation of trees.
Molecular Phylogenetics and Evolution 1992, 1:5358. PubMed Abstract  Publisher Full Text

Goloboff PA: Analyzing Large Data Sets in Reasonable Times: Solutions for Composite Optima.
Cladistics 1999, 15(4):415428. Publisher Full Text

Swofford DL: PAUP*: Phylogenetic analysis using parsimony (*and other methods), Version 4.0b10.
2002. PubMed Abstract  Publisher Full Text

Roshan U, Moret BME, Warnow T, Williams TL: RecIDCM3: A Fast Algorithmic Technique for Reconstructing Large Phylogenetic Trees.

BinindaEmonds O, Sanderson M: Assessment of the accuracy of matrix representation with parsimony analysis supertree construction.
Systematic Biology 2001, 50:565579. PubMed Abstract  Publisher Full Text

Eulenstein O, Chen D, Burleigh JG, FernándezBaca D, Sanderson MJ: Performance of Flip Supertree Construction with a Heuristic Algorithm.
Systematic Biology 2003, 53:299308. Publisher Full Text

Chen D, Eulenstein O, FernándezBaca D, Burleigh JG: Improved Heuristics for MinimumFlip Supertree Construction.

Creevey CJ, McInerney JO: Clann: investigating phylogenetic information through supertree analyses.
Bioinformatics 2005, 21(3):390392. PubMed Abstract  Publisher Full Text

Wilkinson M, Cotton JA: Supertree Methods for Building the Tree of Life: DivideandConquer Approaches to Large Phylogenetic Problems. In Reconstructing the Tree of Life: Taxonomy and Systematics of Species Rich Taxa. Edited by Hodkinson TR, Parnell JAN. CRC Press; 2007:6176.

Robinson DF, Foulds LR: Comparison of phylogenetic trees.
Mathematical Biosciences 1981, 53(12):131147. Publisher Full Text

McMorris FR, Steel MA: The complexity of the median procedure for binary trees.
Proceedings of the International Federation of Classification Societies 1993.

Barthélemy JP, McMorris FR: The median procedure for ntrees.
Journal of Classification 1986, 3:329334. Publisher Full Text

Cotton JA, Wilkinson M: MajorityRule Supertrees.
Systematic Biology 2007, 56:445452. PubMed Abstract  Publisher Full Text

Dong J, FernándezBaca D: Properties of MajorityRule Supertrees.
Syst Biol 2009, 58(3):360367. Publisher Full Text

Day WHE: Optimal algorithms for comparing trees with labeled leaves.
Journal of Classification 1985, 2:728. Publisher Full Text

Pattengale ND, Gottlieb EJ, Moret BME: Efficiently Computing the RobinsonFoulds Metric.
Journal of Computational Biology 2007, 14(6):724735.
[PMID: 17691890]
PubMed Abstract  Publisher Full Text 
Semple C, Steel M: Phylogenetics. Oxford University Press; 2003.

Allen BL, Steel M: Subtree transfer operations and their induced metrics on evolutionary trees.
Annals of Combinatorics 2001, 5:113. Publisher Full Text

Swofford DL, Olsen GJ, Waddel PJ, Hillis DM: Phylogenetic inference. In Molecular Systematics. Edited by Hillis DM, Moritz C, Mable BK. Sunderland, Mass: Sinauer Assoc; 1996:407509.

Bordewich M, Semple C: On the computational complexity of the rooted subtree prune and regraft distance.
Annals of Combinatorics 2004, 8:409423. Publisher Full Text

Bansal MS, Burleigh JG, Eulenstein O, Wehe A: Heuristics for the GeneDuplication Problem: A Θ(n) SpeedUp for the Local Search. In RECOMB, Volume 4453 of Lecture Notes in Computer Science. Edited by Speed TP, Huang H. Springer; 2007:238252. Publisher Full Text

Bansal MS, Eulenstein O: An Ω(n^{2}/log n) SpeedUp of TBR Heuristics for the GeneDuplication Problem.
IEEE/ACM Transactions on Computational Biology and Bioinformatics 2008, 5(4):514524. Publisher Full Text

Lin H, Burleigh JG, Eulenstein O: Triplet supertree heuristics for the tree of life.
BMC Bioinformatics 2009, 10(Suppl 1):S8. PubMed Abstract  BioMed Central Full Text  PubMed Central Full Text

Bender MA, FarachColton M: The LCA Problem Revisited. In LATIN, Volume 1776 of Lecture Notes in Computer Science. Edited by Gonnet GH, Panario D, Viola A. Springer; 2000:8894. Publisher Full Text

Kennedy M, Page R: Seabird supertrees: combining partial estimates of procellariiform phylogeny.
The Auk 2002, 119:88108. Publisher Full Text

Cardillo M, BinindaEmonds ORP, Boakes E, Purvis A: A specieslevel phylogenetic supertree of marsupials.
Journal of Zoology 2004, 264:1131. Publisher Full Text

Beck R, BinindaEmonds O, Cardillo M, Liu FG, Purvis A: A higherlevel MRP supertree of placental mammals.
BMC Evolutionary Biology 2006, 6:93. PubMed Abstract  BioMed Central Full Text  PubMed Central Full Text

Wojciechowski M, Sanderson M, Steele K, Liston A: Molecular phylogeny of the "Temperate Herbaceous Tribes" of Papilionoid legumes: a supertree approach. In Advances in Legume Systematics. Volume 9. Edited by Herendeen P, Bruneau A. Kew: Royal Botanic Gardens; 2000::277298.

Nixon KC: The parsimony ratchet: a new method for rapid parsimony analysis.
Cladistics 1999, 15:407414. Publisher Full Text

Sikes DS, Lewis PO: PAUPRat: PAUP* implementation of the parsimony ratchet.
2001.

Wilkinson M, Pisani D, Cotton JA, Corfe I: Measuring Support and Finding Unsupported Relationships in Supertrees.
Syst Biol 2005, 54(5):823831. PubMed Abstract  Publisher Full Text

BinindaEmonds OR (Ed): Phylogenetic supertrees. Springer Verlag; 2004.

Snir S, Rao S: Using Max Cut to Enhance Rooted Trees Consistency.
IEEE/ACM Trans. Comput. Biology Bioinform 2006, 3(4):323333. Publisher Full Text

Creevey CJ, Fitzpatrick DA, Gayle K Philip RJK, O'Connell MJ, Pentony MM, Travers SA, Wilkinson M, McInerney JO: Does a treelike phylogeny only exist at the tips in the prokaryotes?
Proc Biol Sci 2004, 271(1557):25512558. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Thorley JL, Wilkinson M: A View of Supertree Methods.
Bioconsensus, Volume 61 of DIMACS: Series in Discrete Mathematics and Theoretic Computer Science, Providence, Rhode Island, USA: American Mathematical Society 2003, 185193.

Bansal MS, Eulenstein O, Wehe A: The GeneDuplication Problem: NearLinear Time Algorithms for NNIBased Local Searches.
IEEE/ACM Transactions on Computational Biology and Bioinformatics 2009, 6(2):221231. Publisher Full Text

Bryant D: Building trees, hunting for trees, and comparing trees: Theory and methods in phylogenetic analysis. PhD thesis. Dept. of Mathematics, University of Canterbury; 1997.

Steel M, Rodrigo A: Maximum likelihood supertrees.
Syst. Biol 2008, 57:243250. PubMed Abstract  Publisher Full Text

Kuhner MK, Felsenstein J: A simulation comparison of phylogeny algorithms under equal and unequal evolutionary rates [published erratum appears in Mol Biol Evol 1995 May;12(3):525].
Mol Biol Evol 1994, 11(3):459468. PubMed Abstract  Publisher Full Text