0
BMC Bioinformatics noMed Central
Research I
Efficient genomescale phylogenetic analysis under the duplication
loss and deep coalescence cost models
Mukul S Bansal1'3, J Gordon Burleigh2 and Oliver Eulenstein*3
Addresses: 'School of Computer Science, Tel Aviv University, Tel Aviv 69978, Israel, 2Department of Biology, University of Florida,
Gainesville, FL 32611, USA and 3Department of Computer Science, Iowa State University, Ames, IA 50011, USA
Email: Mukul S Bansal bansal@tau.ac.il; J Gordon Burleigh gburleigh@ufl.edu; Oliver Eulenstein* oeulenst@cs.iastate.edu
* Corresponding author
from The Eighth Asia Pacific Bioinformatics Conference (APBC 2010)
Bangalore, India 1821 January 2010
Published: 18 January 2010
BMC Bioinformatics 2010, I I(Suppl I):S42 doi: 10.1 186/147121051 IS I S42
This article is available from: http://www.biomedcentral.com/14712105/11 /S I/S42
2010 Bansal et al; licensee BioMed Central Ltd.
This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0),
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
Abstract
Background: Genomic data provide a wealth of new information for phylogenetic analysis. Yet
making use of this data requires phylogenetic methods that can efficiently analyze extremely large
data sets and account for processes of gene evolution, such as gene duplication and loss, incomplete
lineage sorting (deep coalescence), or horizontal gene transfer, that cause incongruence among
gene trees. One such approach is gene tree parsimony, which, given a set of gene trees, seeks a
species tree that requires the smallest number of evolutionary events to explain the incongruence
of the gene trees. However, the only existing algorithms for gene tree parsimony under the
duplicationloss or deep coalescence reconciliation cost are prohibitively slow for large datasets.
Results: We describe novel algorithms for SPR and TBR based local search heuristics under the
duplicationloss cost, and we show how they can be adapted for the deep coalescence cost. These
algorithms improve upon the best existing algorithms for these problems by a factor of n, where n
is the number of species in the collection of gene trees. We implemented our new SPR based local
search algorithm for the duplicationloss cost and demonstrate the tremendous improvement in
runtime and scalability it provides compared to existing implementations. We also evaluate the
performance of our algorithm on three largescale genomic data sets.
Conclusion: Our new algorithms enable, for the first time, gene tree parsimony analyses of
thousands of genes from hundreds of taxa using the duplicationloss and deep coalescence
reconciliation costs. Thus, this work expands both the size of data sets and the range of
evolutionary models that can be incorporated into genomescale phylogenetic analyses.
Page 1 of 9
(page number not for citation purposes)
Dpen Access I
BMC Bioinformatics 2010, 11 (Suppl 1):S42
Background
The availability of largescale genomic data sets provides an
unprecedented wealth of information for phylogenetic
analyses. However, genomic data sets also present a
number of unique challenges for phylogenetics. Beyond
the exceptional computational challenges involved in
analyses of thousands of genes from many taxa, complex
patterns of gene evolution within the genome, including
incomplete lineage sorting (deep coalescence), gene
duplications and losses, lateral gene transfer, and recombi
nation, create tremendous heterogeneity in the topology of
gene trees and obscure species relationships [1]. In fact, in
some evolutionary scenarios, trees from the majority of
genes in a genome can differ from the species phylogeny
(e.g., [2]). Thus, methods for incorporating genomic data
into phylogenetic analyses must be both computationally
tractable for extremely large data sets and must account for
heterogeneous processes of gene family evolution. In this
paper, we introduce novel algorithms that enable for the
first time estimates of phylogenetic trees from large
genomic data sets based on gene duplications and losses
as well as incomplete lineage sorting.
One approach for building phylogenetic trees from
genomic data sets is gene tree parsimony (GTP). Given
a collection of gene trees, GTP seeks a species tree that
implies the minimum reconciliation cost, or number of
events that cause conflict among the gene trees. How
ever, available GTP algorithms either lack sufficient
speed to be useful for large data sets or the flexibility to
deal with a wide range of evolutionary processes that
affect gene tree topologies. In particular, the duplication
loss problem [313], which is the GTP problem based on
minimizing the number of gene duplications and losses,
and the deepcoalescence problem [1,1416], which is the
GTP problem based on minimizing the number of deep
coalescences, lack efficient heuristics.
Both the duplicationloss problem and the deepcoales
cence problem are NPhard [15,17]. Therefore, in practice,
these problems are typically approached using local search
heuristics. These heuristics start with some initial candidate
species tree and find a minimum reconciliation cost tree in
its neighborhood. This constitutes one local search step.
The best tree thus found then becomes the starting point
for the next local search step, and so on, until a local
minima is reached. Thus, at each local search step, the
heuristic solves a "local search problem". The time
complexity of this local search problem depends on the
tree edit operation used to define the neighborhood.
Rooted subtree pruning and regrafting (SPR) [18] and
rooted tree bisection and reconnection (TBR) [19] are
two of the most effective and most commonly used tree
edit operations. For example, Page [20] and Maddison
http://www.biomedcentral.com/14712105/11/S 1/S42
and Knowles [14] implemented heuristics based on the
SPR local tree search to estimate the species tree that
minimizes the number of deep coalescences. Similarly,
SPR based local search heuristics were also developed for
the duplicationloss problem [20,21]. However, these
heuristics estimate the reconciliation cost from scratch
for each tree topology that is evaluated, and therefore,
they are only useful for small data sets. Recently, Than
and Nakhleh [16] developed an integer linear program
ming formulation and a dynamic programming algo
rithm to provide exact solutions for the deep coalescence
problem. Although these exact algorithms can analyze
data sets with hundreds of gene trees, they are limited to
a very small number of taxa. Furthermore, while there do
exist fast heuristics based on SPR [22] and TBR [23] local
searches for GTP, these are based on minimizing the
gene duplication cost only.
Several methods exist for inferring species trees from
collections of conflicting genes in a probabilistic frame
work, but, these are similarly limited. Liu and Pearl [24]
and Kubatko et al. [25] discuss likelihoodbased
approaches to estimate a species tree from gene trees
based on a coalescent process. These methods, along
with An6 et al.'s [26] Bayesian approach to estimating
concordance among gene trees, require gene trees with a
single gene per taxon, and they have only been tested on
small data sets. In addition, probabilistic models of gene
treespecies tree reconciliation that incorporate duplica
tion and loss also exist [27,28]. However, these models
are computationally complex, and they have not been
incorporated in any tree search heuristics.
Our contributions
A lack of efficient heuristics has limited the use of the
GTP approach for phylogenetic analyses of largescale
genomic data sets, and there are no options for such GTP
analyses based on the duplicationloss and deep
coalescence problems. In this paper, we address this
issue by presenting efficient, novel algorithms for SPR
and TBR based local searches for both of these problems.
Let us assume, for convenience, that the size of the k
given gene trees differs by a constant factor from the size
of the resulting species tree. The currently best known
(naive) solutions for the SPR and TBR local search
problems, for the duplicationloss as well as the deep
coalescence problem, require O(kn3) and O(kn4) time
respectively, where n is the size of the resulting species
tree. Our new algorithms solve these SPR and TBR local
search problems in O(kn2) and O(kn3) time respectively.
Consequently, our algorithms provide a speedup of a
factor of n over the best known SPR and TBR local search
algorithms (like the ones implemented in [20,21]) for
the duplicationloss and deepcoalescence problems.
Page 2 of 9
(page number not for citation purposes)
BMC Bioinformatics 2010, 11 (Suppl 1):S42
This enables, for the first time, GTP analyses with
hundreds of taxa and thousands of genes based on
gene duplications and losses or incomplete lineage
sorting.
We first develop our algorithms in the context of the
duplicationloss problem, and then show how to apply
them to the deepcoalescence problem. Then, we
demonstrate the improvement in runtime and scalability
provided by our algorithms over the best current
solutions by using an implementation of our algorithm
for the SPR local search.
Methods
Basic notation and preliminaries
Given a rooted tree T, we denote its node set, edge set,
and leaf set by V (T), E(T), and Le(T) respectively. The
root node of T is denoted by rt(T). Given a node v e V
(T), we denote its parent by pat (v), its set of children by
ChT (v), and the subtree of T rooted at v by T,. If two
nodes in T have the same parent, they are called siblings.
The set of internal nodes of T, denoted I(T), is defined to
be V (T)\Le(T). 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 least common ancestor of a nonempty subset L
 V (T) in tree T, denoted as IcaT (L), is the unique
smallest upper bound of L under <.
Given x, y e V (T), x >T y denotes the unique path from x
to y in T. We denote by dT (x, y) the number of edges on
the path x >T y. T is fully binary if every node has either
zero or two children. Throughout this paper, the term
tree refers to a rooted fully binary tree. Given T and a set
L c Le(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 DuplicationLoss problem
A species tree is a tree that depicts the evolutionary
relationships of a set of species. Given a gene family for a
set of species, a gene tree is a tree that depicts the
evolutionary relationships among the sequences encod
ing only that gene family in the given set of species.
Thus, the nodes in a gene tree represent genes. We
assume that each leaf of the gene trees is labeled with the
species from which that gene was sampled. In order to
compare a gene tree G with a species tree S, we require a
mapping from each gene g e V (G) to the most recent
species in S that could have contained g.
Definition 1 (Mapping). The leafmapping c, s : Le(G)
> Le(S) maps a leaf node g e Le(G) to that unique leaf node
http://www.biomedcentral.com/14712105/11/S 1/S42
s e Le(S) which has the same label as g. The extension _c, s :
V (G) > V (S) of LG, s is the mapping defined by &cG, s(g)
= lca(L, s(Le(G,))).
For any node s e V (S), we use MGls(s) to denote the set
of nodes in G that map to node s e V (S) under the
mapping .&G, s.
Definition 2 (Comparability). Given trees G and S, we say
that G is comparable to S if, for each g e Le(G), the leaf
mapping Lc, s(g) is well defined. A set of gene trees G is
comparable to S if each gene tree in G is comparable to S.
Throughout this paper we use the following terminol
ogy: G is a set of gene trees that is comparable to a
species tree S, and G e G.
Definition 3 (Duplication). A node g e I(G) is a (gene)
duplication if Ic, s(g) e _&c, s(Ch(g)) and we define Dup
(G, S) = {g e I(G): g is a duplication}.
Following [8], we define the number of losses as follows.
Definition 4 (Losses). The number of losses Loss(G, S, g) at
a node g e I(G), is defined to be:
0, if AG, S' (g) = G, s' (g') V g' e Ch(g), and
Zg Ch(gs) ds, (G, s' (g), cG, s' (g')) 1, otherwise;
where S' = S[Le(G)]. We define Loss(G, S) = sEgc(G) Loss(G,
S, g) to be the number of losses in G.
Under the duplicationloss model, the reconciliation
cost of G with S is simply the duplicationloss cost; i.e.,
the number of duplications and losses.
Definition 5 (Reconciliation cost). We define reconcilia
tion costs for gene and species trees as follows:
1. A(G, S) = I Dup(G, S) + Loss(G, S) is the
reconciliation cost from G to S.
2. A(G, S) =YGQ G A(G, S) is the reconciliation cost
from G to S.
3. Let T be the set of species trees that are comparable
with G. We define A() = minsT A(G, S) to be the
reconciliation cost of g .
Problem 1 (DuplicationLoss). Given a set G of gene trees,
the DuplicationLoss problem is to find a species tree S*
comparable with G, such that A ( G, S*) = A( G).
Local search problems
Here we first provide the definition of an SPR edit
operation [18] and then formulate the related local
search problems. The definition and associated local
Page 3 of 9
(page number not for citation purposes)
BMC Bioinformatics 2010, 11 (Suppl 1):S42
search problems for the TBR edit operation are con
sidered later.
For technical reasons, before we can define the SPR
operation, we need the following definition.
Definition 6 (Planted tree). Given a tree T, the planted
tree Q((T) is the tree obtained by adding a root edge {p, rt
(T)}, where p 4 V (T), to T.
Definition 7 (SPR operation). Let T be a tree, e = {u, v} e
E(T), where u = pa(v), and X, Y be the connected components
that are obtained by removing edge e from T such that v e X
and u e Y. We define SPRT (v, y) for y e Y to be the tree that
is obtained from uI(T) by first removing edge e, and then
adjoining a new edge f between v and Y as follows:
1. Create a new node y' that subdivides the edge {pa(y),
y}.
2. Add edge f between nodes v and y'.
3. Suppress the node u, and rename y' as u.
4. Contract the root edge.
We say that the tree SPRT (v, y) is obtained from T by a
subtree prune and regraft (SPR) operation that prunes
subtree T, and regrafts it above node y.
Notation. We define the following:
1. SPRT (v) =U, y{SPRT (v, y)}
2. SPRT = U (., v)EE(T) SPRT (v)
Throughout the remainder of this manuscript, S denotes
a species tree such that Le(S) = UGeUgLe(G)LGS(g),
and v is a nonroot node in V (S).
We now define the relevant local search problems based
on the SPR operation.
Problem 2 (SPRScoring (SPRS)).
Given G and S, find a tree T *E SPRs such that A ( G, T *) =
minT.eSRs A (g, T).
Our goal is to solve the SPRS problem efficiently. To
that end, we first define a restricted version of the SPRS
problem, called the SPRRestricted Scoring problem.
Problem 3 (SPRRestricted Scoring (SPRRS)).
Given G, S, and v, find a tree T *E SPRs(v) such that A ( G,
T *) = minTSPRs(v) A(G, T).
Let n =  Le(S), m = Le(S) + Le(G) and k = G and
let us assume, for convenience, that all G e g have
approximately the same size. In the following, we show
http://www.biomedcentral.com/14712105/11/S 1/S42
how to solve the SPRRS problem in O(km) time. Since
SPRs = U {pa(v), v} e E(S) SPRs(v), it is easy to see that the
SPRS problem can be solved by solving the SPRRS
problem 0(n) times. This yields an O(kmn) time
algorithm for the SPRS problem. Later, we show that
the local search problem corresponding to the TBR
operation reduces to solving 0(n2) SPRRS problems;
which yields an 0(kmn2) time algorithm for the TBR
local search problem. In the interest of brevity, all
lemmas and theorems in this paper appear with proofs
omitted; however, all proofs are available in [29].
Solving the SPRRS problem
Throughout this section, we limit our attention to one
gene tree G; in particular, we show how to solve the SPR
RS problem for G in 0(m) time. Our algorithm extends
trivially to solve the SPRRS problem on the set of gene
trees G in O(km) time. For simplicity, we will assume
that Le(G) = Le(S). Indeed, if Le(G) # Le(S) then we can
simply set the species tree to be S[Le(G)]; this takes 0(n)
time and, consequently, does not affect the time
complexity of our algorithm.
In order to solve the SPRRS problem for G, it is
sufficient to compute the values I Dup(G, S') and Loss(G,
S') for each S' e SPRs(v). Bansal et al. [22] showed how
to compute the value I Dup(G, S') for each S' e SPRs(v),
in 0(m) time. Losses, however, behave in a very different
and much more complex manner compared to gene
duplications and it has remained unclear if their
computation could be similarly optimized. In this
paper we show that it is indeed possible to compute
the value Loss(G, S') for each S' e SPRs(v) in 0(m) time
as well. Altogether, this implies that the SPRRS problem
for G can be solved in 0(m) time. Next, we introduce
some of the basic structural properties that are helpful in
the current setting.
Basic structural properties
Consider the tree Ns' v = SPRs(v, rt(S)). Observe that, since
SPRNs, (v) = SPRs(v), solving the SPRRS problem on
instance ({ G }, S, v) is equivalent to solving it on the instance
({G }, Ns' v, v). Thus, in the remainder of this section, we will
work with tree Ns' v instead of tree S; the reason for this
choice becomes clear in light of Lemmas 3 and 4.
Since S and v are fixed in the current context, we will, in
the interest of clarity, abbreviate Ns5 simply to N.
Similarly, in the remainder of this section, we abbreviate
&c, T to &Tr, for any species tree T.
Throughout the remainder of this work, let u denote the
sibling of v in N. We color the nodes ofN asfollows: (i)
All nodes in the subtree Nv are colored red, (ii) the root
Page 4 of 9
(page number not for citation purposes)
BMC Bioinformatics 2010, 11 (Suppl 1):S42
node of N is colored blue, and (iii) all the remaining
nodes, i.e. all nodes in N,,, are colored green. Corre
spondingly, we color the nodes of G by assigning to each
g e V (G) the color of the node ,' (.,1.
Definition 8 (F). We define F to be the tree obtained from G
by removing all red nodes (along with any edges incident on
these red nodes). Observe that while F must be binary, it
might not be fully binary.
The significance of the tree F stems from the following
two Lemmas which, together, completely characterize
the mappings from nodes in V (G) for each S' e SPRs(v).
This characterization is the basis of Lemmas 3 through 8.
Lemma 1. Given G and N, if g e V (G) is either red or green,
then .isM(g) = (:1 for all S' e SPRN (v).
Lemma 2. Given G and N, if g e V (G) is a blue node, then
_&s, (g) = lcas, (v, &F, N (g)) for any S' e SPRN(v).
Characterizing losses
To solve the SPRRS problem efficiently we rely on the
following six lemmas, which make it possible to
efficiently infer the value of Loss(G, S', g) for any S' e
SPRN (v) and any g e V (G).
Consider any g e I(G), and let g' and g" be its two
children. Let a = ,' (.,', b = _&N (g') and c = _&N (g").
Without loss of generality, node g must correspond to
one of the following six categories: 1) g is red, 2) g is
green, 3) g, g', and g" are all blue, 4) g and g' are blue,
and g" is green, 5) g and g' are blue, and g" is red, or, 6) g
is blue, g' is red, and g" is green.
Lemmas 3 through 8 characterize the behavior of the loss
cost Loss(G, S', g), for each S' e SPRN (v), for each of
these six cases. At this point, it would help to observe
that SPRN (v) = {SPRN (v, s): s e V (N,,)}.
Lemma 3. If g is red then Loss(G, S', g) = Loss(G, N, g) for
all S' e SPRN (v).
Lemma 4. If g is green then Loss(G, S', g) = Loss(G, N, g) + 1
ifS' = SPRN (v, x) where b
(G, S', g) = Loss(G, N, g) otherwise.
Lemma 5. Let g, g' and g" all be blue nodes, x e V (N,,), and
let a' = .r4, N (g), b' = N.&, N (g') and c' = .&r, N (g").
1. If S' = SPRN (v, x) where x MN a', then Loss(G, S', g) =
Loss(G, N, g).
2. IfS' = SPRN (v, x) where x
(x)), then,
http://www.biomedcentral.com/14712105/11/S 1/S42
(a) Loss(G, S', g) = Loss(G, S", g) + 1 ifb'
(b) Loss(G, S', g) = Loss(G, S", g) otherwise.
Lemma 6. Let g and g' be blue nodes and g" be a green node,
x e V (N,)\{u}, and let a' = &, N (g), b' = N&F, N (g') and
c'= &, N (g").
1. If S' = SPRN (v, x) where x MN a', and S" = SPRN (v, pa
(x)), then,
(a) Loss(G, S', g) = Loss(G, S", g) 1 if a'
(b) Loss(G, S', g) = Loss(G, S", g) 1 if a' < N pa(x)
x is not such that a'
(c) Loss(G, S', g) = Loss(G, S", g) otherwise.
2. Let S' = SPRN (v, x) where x
(x)).
(a) If a' # b' and b" denotes the child of a' along the path
a' > N b', then,
i. Loss(G, SPRN (v, b"), g) = Loss(G, SPRN (v, a'), g) 2 if a'
> c'. And,
Loss(G, SPRN (v, b"), g) = Loss(G, SPRN (v, a'), g) if a' =
c,
ii. Loss(G, S', g) = Loss(G, S", g) + 1 if b' < N x
iii. Loss(G, S', g) = Loss(G, S", g) if x is such that x e V (Nb..")
but not such that b' < N x
iv. Loss(G, S', g) = Loss(G, SPRN (v, a'), g) if c'
and,
v. Loss(G, S', g) = Loss(G, SPRN (v, a'), g) 1 otherwise.
(b) If a' = b', then,
i. Loss(G, S', g) = Loss(G, SPRN (v, a'), g) if c' < N x
and,
ii. Loss(G, S', g) = Loss(G, SPRN (v, a'), g) 1 otherwise.
Lemma 7. Let g and g' be blue nodes and c be a red node, x e
V (Nj), and let a' = .YF, N(g).
1. If S' = SPRN (v, x) where x
(x)), then Loss(G, S', g) = Loss(G, S", g) + 1.
2. If S' = SPRN (v, x) where x MN a', then,
(a) Loss(G, S', g) = Loss(G, N, g) if a'
(b) Loss(G, S', g) = Loss(G, S", g) + 1 for S" = SPRN (v, pa
(x)) otherwise.
Lemma 8. Let g be blue, g' be red, and g" be green. Let x e V
(N,)\{u} and a' = .F, N (g).
1. If S' = SPRN (v, x) where x MN a', and S" = SPRN (v, pa
(x)), then,
(a) Loss(G, S', g) = Loss(G, S", g) 1 if a' < N x
(b) Loss(G, S', g) = Loss(G, S", g) if a' < N pa(x)
not such that a'
(c) Loss(G, S', g) = Loss(G, S", g) + 1 otherwise.
2. If S' = SPRN (v, x) where x
(x)), then,
(a) Loss(G, S', g) = Loss(G, SPRN (v, a'), g) + 2 if x e ChN
(a'), and,
(b) Loss(G, S', g) = Loss(G, S", g) + 1 otherwise.
Page 5 of 9
(page number not for citation purposes)
BMC Bioinformatics 2010, 11 (Suppl 1):S42
The algorithm
Observe that SPRN (v) = {SPRN (v, s): s e V (N,)}.
Therefore, the goal of our algorithm is to compute at
each node s e V (N,) the value Loss(G, S'), where S' =
SPRN (v, s). To do this efficiently, we rely on the
characterization of losses given in Lemmas 3 through 8.
The first step of the algorithm is to compute the value
Loss(G, N). This "loss value" is assigned to the node u. To
compute the loss value for the rest of the nodes our
algorithm makes use of six different types of counters at
each node in N,; we refer to these counters as counteri,
for i e { 1..., 6}. The reason for using these six counters is
that the behavior of the loss values can be explained by
using six types of patterns (captured by the six counters).
These counters make it possible to efficiently compute
the difference between the values Loss(G, N) and Loss(G,
S'), where S' = SPRN (v, s), for each s e V (N,). Next, we
describe each of these six counters; throughout our
description, s represents some node in N,.
counter1 If the value of counter1 is x at node s then this
implies that the tree SPRN (v, s) incurs x additional losses
over the value Loss(G, N).
counter2 If the value of counter2 is x at node s, then
this implies that for each t
an additional x losses over Loss(G, N).
counter3 If the value of counter3 is x at node s, then
this implies that for each t< N s the tree SPRN (v, t) loses x
losses from Loss(G, N).
counter4 If the value of counter4 is x at node s, then this
implies that for each t < N s the tree SPRN (v, t) incurs a, x
additional losses over Loss(G, N, where a, = dN (pa(s), t).
counter5 If the value of counter5 is x at node s, then it
is equivalent to incrementing counter4 at the sibling of
each node on the path u > N s, except at u, by x.
counter6 If the value of counter6 is x at node s, then it
is equivalent to incrementing counter4 at both children
(if they exist) of the sibling of each node along the path u
> N s, except u, and incrementing counter3 at each node
along the path u > N s, except at u, by x.
In the remainder of this section we first show how to
compute the values of these counters, and then the final
loss values, at each node in N,.
Computing the counters
We now describe how the values of the six counters are
computed. Initially, each counter at each node in N, is
http://www.biomedcentral.com/14712105/11/S 1/S42
set to 0. Consider any g e I(G), and let g' and g" be its
two children. Recall that node g must fall under one of
the following six categories: 1) g is red, 2) g is green, 3) g,
g', and g" are all blue, 4) g and g are blue, and g" is
green, 5) g and g' are blue, and g" is red, or, 6) g is blue, g'
is red, and g" is green.
Let a = .&N (g), b = .&N (g') and c = .&N (g"). Also,
whenever properly defined, let a' = r, N (g), b' = &r, N
(g') and c'= &r, N (g"). Based on Lemmas 3 through 8,
we now study how the six counters can be updated so as
to capture the behavior of losses in each of these cases.
Case 1. By Lemma 3, we do nothing in this case.
Case 2. Based on Lemma 4, the contribution of any node
g that satisfies the condition of case 2 can be captured by
incrementing the value of counter1 by one at each node
on paths a > N b and a > N b, except at node a.
Case 3. From Lemma 5 it follows that in this case the
contribution of g to the loss value changes in a way that
is captured by incrementing counter2 by 1, at each
node, except a', on the paths a' > N b' and a'> N c'.
Case 4. According to Lemma 6, if N, is regrafted on an
edge of N, that is not in Na,, then the contribution of g to
the loss cost is captured by incrementing counter3 by 1
at each node except u along the path u >N a', and at their
siblings. If N, is regrafted on an edge of N, that is in Na,
then there are two possible cases:
a' ; b': Recall that b" represents the child of a' along the
path a' > N b'. Here, the contribution ofg to the loss cost
is captured by (i) incrementing counter3 by two at node
b", (ii) incrementing counter2 by one at each node
except b" along the path b" >N b", (iii) incrementing
counter3 by one at the sibling of b", and (iv)
incrementing counter1 by one at each node except a'
on the path a' > N c'.
a' = b': In this case, the contribution ofg to the loss cost is
captured by (i) incrementing counter3 by one at both
children of a', and (ii) incrementing counter1 by one at
each node except a' on the path a' > N c'.
Case 5. By Lemma 7, for this case, the change in the loss
contribution of g is captured by incrementing counter5
by 1 at node a', and by incrementing counter4 by 1 at
both children of a' in N.
Case 6. By Lemma 8, for this case, the change in the loss
contribution of g is captured by incrementing counter6
by 1 at node a', and by incrementing counter4 and
counter2 by 1 each at both children of a' in N.
Page 6 of 9
(page number not for citation purposes)
BMC Bioinformatics 2010, 11 (Suppl 1):S42
Computing the final loss values
Our algorithm considers each internal node of gene tree
G, one at a time, and updates the relevant counters at the
relevant nodes in Nu, as shown in the previous
subsection. Then, based on the counters, it computes,
at each node s e V (N,) the value a (s) = Loss(G, S') Loss
(G, N), where S' = SPRN (v, s); this can be achieved by
performing a constant number of pre and postorder
traversals of N,. A final traversal of the tree now allows
us to compute the value Loss(G, S') = a(s) + Loss(G, N) at
each s e V (N,). Due to space limitations, a more
complete description of the algorithm is omitted from
this manuscript (but is available in [29]).
For simplicity in stating the time complexity of our
algorithm, we assume that all G e g have approximately
the same size. Recall that n = Le(S) m =  Le(S) + I Le(G)
and k = G I.
Lemma 9. The SPRRS problem on the input instance ({G},
S, v) can be solved in 0(m) time.
Remark on Lemma 9: Observe that in cases 2, 3 and 4
(from the previous subsection), handling each g might
require updating the counters at 0(n) nodes, yielding a
total time complexity of O0(nm) for these cases. However,
it is still possible to obtain the 0(m) time bound.
Thus, we have the following theorem.
Theorem 1. The SPRRS and SPRS problems can be solved
in O(km) and O(kmn) time respectively.
The time complexity of the best known (naive) solution
for the SPRS problem is 0 (kmn2). Our algorithm
improves on this by a factor of n.
Speedingup the TBR local search problem
Intuitively, a (rooted) TBR operation may be viewed as
being like an SPR operation except that the TBR
operation allows the pruned subtree to be arbitrarily
rerooted before being regrafted. The TBRS problem is
defined analogously to the SPRS problem.
Observe that there are (n) different ways to select a subtree
of S to be pruned. Furthermore, there are 0(n) different
ways to reroot the pruned subtree. The idea is to directly use
the solution to the SPRRS problem to compute the
duplication and loss costs for the 0(n)cardinality subset
of TBRs defined by any fixed pruned subtree and its fixed
rooting. This yields the following theorem.
Theorem 2. The TBRS problem can be solved in 0(kmn2)
time.
http://www.biomedcentral.com/14712105/11/S 1/S42
This improves on the best known solution for the TBRS
problem by a factor of n.
The deep coalescence cost model
Our algorithms for efficient SPR and TBR local searches
for the duplicationloss model apply directly to the
corresponding SPR and TBR local search problems for
the deep coalescence model. This can be achieved using
any of the following two methods. The first method is to
make use of the result of Zhang [15] who showed
showed that if G is a uniquely leaflabeled gene tree and
S is a species tree such that Le(G) = Le(S), then the deep
coalescence cost of G and S is equal to Loss(G, S) 2 Dup
(G, S). Thus, our algorithms imply that we can compute
the deep coalescence cost of each tree in the SPR (resp.
TBR) neighborhood of S in 0(n2) (resp. 0(n3)) time. The
second method is to use the algorithm for computing
losses, presented in this paper, and slightly modifying it
to directly compute the deep coalescence cost. Owing to
the similarity in the definition of the loss cost and the
deep coalescence cost, this can be done in a straightfor
ward manner (details omitted for brevity). Overall our
algorithms yield speedups of a factor of n over the
fastest current approaches for SPR and TBR local searches
for the deep coalescence cost model.
Results
To evaluate the performance of our novel local search
algorithms, we implemented our algorithm for the SPRS
problem as part of a standard search heuristic for the
duplicationloss problem. We refer to our program as
DupLoss. We first evaluated its performance on randomly
generated gene trees. These data sets represent extreme
examples of incongruence among gene trees, and thus,
they provide a challenging way to test the runtime
performance of a gene tree reconciliation method. The
input gene trees for each run consisted of 20 trees, each
with the same set of taxa and with random binary
topologies and random assignment of leaf labels. We
conducted runs with 50, 100, 200, 400, and 1000 taxa in
each gene tree. All analyses were performed on a 3 Ghz
Intel Pentium 4 CPU based PC with Windows XP
operating system. We compared the performance of
DupLoss with the program GeneTree [20], which, like
Mesquite [21], implements similar local search heuristics
based on the best known (naive) algorithm for the SPRS
problem. Our implementation shows a tremendous
improvement in runtime and scalability compared to
GeneTree (Table 1). For example, on the 200 taxon data
set, our implementation finished in less than five
minutes, while GeneTree ran for almost six days
(Table 1). Furthermore, we could not run GeneTree
when the input trees had more than 200 taxa.
Page 7 of 9
(page number not for citation purposes)
BMC Bioinformatics 2010, 11 (Suppl 1):S42
Table I: GeneTree vs. DupLoss. Comparison of the runtimes of
GeneTree and DupLoss on the same randomly generated
datasets. Times are given in days(d), hours(h), minutes(m), and
seconds(s).
Taxa size GeneTree DupLoss
50 II m:42 s 5 s
100 3 h:57 m 33 s
200 5 d:19 h:49 m 4 m:24 s
400 43 m:08 s
1000 19 h:27 m
We also tested the performance of DupLoss using several
empirical data sets. First, we ran our implementation on
the 8taxon, 106gene yeast data set of Rokas et al. [30],
and the 8taxon, 268gene Apicomplexan data set of Kuo
et al. [31]. For the yeast data set, we made the gene trees
using maximum likelihood implemented in RAxML [32],
and for the Apicomplexan data set, we used the gene
trees included as supplemental data [31]. Tree searches
for both data sets finished within one second, and the
topologies were consistent with the unrooted species
trees presented in each study. Finally, we ran our
implementation on a plant data set consisting of 18,
896 gene trees from 136 taxa, previously used in a gene
tree reconciliation analysis based on duplications only
[33]. Although this is among the largest data sets used in
a gene tree reconciliation analysis, our heuristic finished
in approximately 24 hours, and the resulting species tree
was consistent with the general consensus of plant
relationships. DupLoss is freely available upon request.
Discussion
The novel algorithms presented in this paper enable GTP
analyses that incorporate gene duplications and losses as
well as deep coalescence on a scale that is impossible
with previous implementations. Whereas most phyloge
netic analyses only use data from putative orthologs, the
duplicationloss model allows one to incorporate data
from large gene families into phylogenetic analysis,
and it does not depend on the accuracy of orthology
estimates. Incorporating deep coalescence events may be
critical when there is a history of rapid speciation, and
these evolutionary scenarios represent many of the most
difficult, and interesting, phylogenetic problems.
Previous advances in GTP heuristics for the duplication
cost problem [22,23,34,35], that is, calculating the
species tree based on the number of duplications only
without considering losses, have enabled promising
phylogenetic analyses of extremely large data sets (e.g.,
136 taxa and 18, 896 genes in [33]). With partial
sequence data, it is difficult, if not impossible, to
distinguish losses from missing sequence data. Thus, it
has been argued that the duplication only reconciliation
http://www.biomedcentral.com/14712105/11/S 1/S42
cost is more appropriate than the duplicationloss cost
when analyzing incomplete data sets (e.g., [36]).
However, with the rapid accumulation of complete
genome sequences, the duplicationloss problem pro
vides a more complete model of evolution than the
duplication problem. Similarly, there has been much
recent interest in the deep coalescence problem (e.g.,
[14]), but it has not been applied to large data sets due
to a lack of efficient heuristics.
While GTP has been effective on small data sets (see review
in [36]), its performance in general is relatively unchar
acterized, largely due to the lack of fast implementations.
The parsimony approach minimizes the number of
evolutionary events that are counted, and therefore, it
may not be appropriate when genes exhibit high rates of
duplication and loss or deep coalescence events. In such
cases, inferring the species tree based on a likelihood
model of gene evolution may be more appropriate. Still,
likelihood methods are often computationally burden
some, and the computational difficulties are compounded
by the extremely large data sets that are an inherent part of
genomescale analysis. The algorithms in this paper
provide a pragmatic solution to the problem of addressing
complex processes of evolution on enormous data sets in a
phylogenetic analysis.
Conclusion
The abundance of new genomic sequence data presents
new challenges for phylogenetic inference. Genomescale
phylogenetic analyses must account for complex evolu
tionary processes, such as gene duplication and loss,
incomplete lineage sorting (deep coalescence), or hor
izontal gene transfer, that can produce conflict among
gene trees, and they must be computationally feasible for
enormous data sets. Our new algorithms for fast local
tree search for inferring species trees under the duplica
tionloss and deep coalescence reconciliation costs
expand both the size of the data sets and the range of
evolutionary models that can be incorporated into
genomescale phylogenetic analyses.
Competing interests
The authors declare that they have no competing
interests.
Authors' contributions
MSB was responsible for algorithm design and program
implementation, and wrote major parts of the paper. JGB
performed the experimental evaluation and the analysis of
the results, and contributed to the writing of the manu
script. OE supervised the project and contributed to the
writing of the paper. All authors read and approved the
final manuscript.
Page 8 of 9
(page number not for citation purposes)
BMC Bioinformatics 2010, 11 (Suppl 1):S42
Acknowledgements
This work was supported in part by NSF grants 0334832 and 0830012. In
addition, MSB was supported in part by a postdoctoral fellowship from the
Edmond J. Safra Bioinformatics program at TelAviv university.
This article has been published as part of BMC Bioinformatics Volume I I
Supplement I, 2010: Selected articles from the Eighth AsiaPacific
Bioinformatics Conference (APBC 2010). The full contents of the
supplement are available online at http://www.biomedcentral.com/1471
2105/ Il?issue=S I.
References
1. Maddison WP: Gene Trees in Species Trees. Systematic Biology
1997, 46:523536.
2. Degnan JH and Rosenberg NA: Discordance of Species Trees
with Their Most Likely Gene Trees. PLoS Genetics 2006, 2(5):
e68.
3. Goodman M, Czelusniak J, Moore GW, RomeroHerrera AE and
Matsuda G: Fitting the gene lineage into its species lineage. A
parsimony strategy illustrated by cladograms constructed
from globin sequences. Systematic Zoology 1979, 28:132163.
4. Page RDM: Maps between trees and cladistic analysis of
historical associations among genes, organisms, and areas.
Systematic Biology 1994, 43:5877.
5. Guig6 R, Muchnik I and Smith TF: Reconstruction of Ancient
Molecular Phylogeny. Molecular Phylogenetics and Evolution 1996, 6
(2):189213.
6. Mirkin B, Muchnik I and Smith TF: A Biologically Consistent
Model for Comparing Molecular Phylogenies. journal of
Computational Biology 1995, 2(4):493507.
7. Eulenstein 0 and Vingron M: On the equivalence of two tree
mapping measures. Discrete Applied Mathematics 1998,
88:101126.
8. Hallett MT and Lagergren J: New algorithms for the duplication
loss model. RECOMB 2000, 138146.
9. Bonizzoni P, Vedova GD and Dondi R: Reconciling a gene tree to
a species tree under the duplication cost model. Theor Comput
Sci 2005, 347(12):3653.
10. G6recki P and Tiuryn J: DLStrees: A model of evolutionary
scenarios. Theor Comput Sci 2006, 359(13):378399.
II. Durand D, Halld6rsson BV and Vernot B: A Hybrid Micro
Macroevolutionary Approach to Gene Tree Reconstruction.
journal of Computational Biology 2006, 13(2):320335.
12. Chauve C, Doyon JP and EIMabrouk N: Gene Family Evolution
by Duplication, Speciation, and Loss. journal of Computational
Biology 2008, 15(8): 10431062.
13. Chauve C and EIMabrouk N: New Perspectives on Gene Family
Evolution: Losses in Reconciliation and a Link with Super
trees. RECOMB 2009, 4658.
14. Maddison WP and Knowles LL: Inferring Phylogeny Despite
Incomplete Lineage Sorting. Systematic Biology 2006, 55:2130.
15. Zhang L: Inferring a Species Tree from Gene Trees under the
Deep Coalescence Cost. RECOMB 2000, 192193.
16. Than C and Nakhleh L: Species tree inference by minimizing
deep coalescences. PLoS Computational Biology 2009, 5(9):
el 000501.
17. Ma B, Li M and Zhang L: From Gene Trees to Species Trees.
SIAM j Comput 2000, 30(3):729752.
18. Bordewich M and Semple C: On the computational complexity
of the rooted subtree prune and regraft distance. Annals of
Combinatorics 2004, 8:409423.
19. Chen D, Eulenstein 0, FernindezBaca D and Burleigh JG:
Improved Heuristics for MinimumFlip Supertree Construc
tion. Evolutionary Bioinformatics 2006, 2:347356.
20. Page RDM: comparing gene and species phylogenies using
reconciled trees. Bioinformatics 1998, 14(9):819820.
21. Maddison WP and Maddison D: Mesquite: a modular system for
evolutionary analysis. Version 2.6.2009 http://mesquiteproiect.
orR.
22. Bansal MS, Burleigh JG, Eulenstein 0 and Wehe A: Heuristics for
the GeneDuplication Problem: A 0(n) SpeedUp for the
Local Search. RECOMB 2007, 238252.
23. Bansal MS and Eulenstein 0: An n(n2/log n) SpeedUp of TBR
Heuristics for the GeneDuplication Problem. IEEE/ACM
Transactions on Computational Biology and Bioinformatics 2008, 5
(4):514524.
http://www.biomedcentral.com/14712105/11/S 1/S42
24. Liu L and Pearl DK: Species Trees from Gene Trees:
Reconstructing Bayesian Posterior Distributions of a Spe
cies Phylogeny Using Estimated Gene Tree Distributions.
Systematic Biology 2007, 56(3):504514.
25. Kubatko LS, Carstens BC and Knowles LL: STEM: species tree
estimation using maximum likelihood for gene trees under
coalescence. Bioinformatics 2009, 25(7):971973.
26. Ane C, Larget B, Baum DA, Smith SD and Rokas A: Bayesian
Estimation of Concordance Among Gene Trees. Mol Biol Evol
2007, 24(7):1575.
27. Arvestad L, Berglund AC, Lagergren J and Sennblad B: Bayesian
gene/species tree reconciliation and orthology analysis using
MCMC. ISMB (Supplement of Bioinformatics) 2003, 715.
28. Akerborg 0, Sennblad B, Arvestad L and Lagergren J: Simulta
neous Bayesian gene tree reconstruction and reconciliation
analysis. Proceedings of the National Academy of Sciences 2009, 106
(14):57145719.
29. Bansal MS: Algorithms for efficient phylogenetic tree con
struction. PhD thesis Iowa State Univ; 2009.
30. Rokas A, Williams BL, King N and Carroll SB: Genomescale
approaches to resolving incongruence in molecular phylo
genies. Nature 2003, 425:798804.
31. Kuo CH, Wares JP and Kissinger JC: The Apicomplexan Whole
Genome Phylogeny: An Analysis of Incongruence among
Gene Trees. Mol Biol Evol 2008, 25(12):26892698.
32. Stamatakis A: RAxMLVIHPC: maximum likelihoodbased
phylogenetic analyses with thousands of taxa and mixed
models. Bioinformatics 2006, 22(21):26882690.
33. Burleigh JG, Bansal MS, Eulenstein 0, Hartmann S, Wehe A and
Vision TJ: Genomescale phylogenetics: inferring the plant
tree of life from 18,896 discordant gene trees. Systematic
Biology in press.
34. Bansal MS, Eulenstein 0 and Wehe A: The GeneDuplication
Problem: NearLinear Time Algorithms for NNIBased
Local Searches. IEEE/ACM Transactions on Computational Biology
and Bioinformatics 2009, 6(2):22123 1.
35. Wehe A, Bansal MS, Burleigh JG and Eulenstein 0: DupTree: a
program for largescale phylogenetic analyses using gene
tree parsimony. Bioinformatics 2008, 24(13):.
36. Cotton JA and Page RDM: Tangled tales from multiple
markers: reconciling conflict between phylogenies to build
molecular supertrees. Phylogenetic Supertrees: Combining Informa
tion to Reveal the Tree of Life SpringerVerlag: BinindaEmonds ORP
2004, 107125.
Page 9 of 9
(page number not for citation purposes)
Publish with BioMed Central and every
scientist can read your work free of charge
"BioMed Central will be the most significant development for
disseminating the results of biomedical research in our lifetime."
Sir Paul Nurse, Cancer Research UK
Your research papers will be:
available free of charge to the entire biomedical community
peer reviewed and published immediately upon acceptance
cited in PubMed and archived on PubMed Central
yours you keep the copyright
Submit your manuscript here: BioMedcentral
http://www biomedcentral.com/info/publishingadv.asp
