Reconciliationbased detection of coevolving gene families
 Yaoban Chan^{1},
 Vincent Ranwez^{2, 3} and
 Celine Scornavacca^{1, 3}Email author
DOI: 10.1186/1471210514332
© Chan et al.; licensee BioMed Central Ltd. 2013
Received: 8 July 2013
Accepted: 13 November 2013
Published: 20 November 2013
Abstract
Background
Genes located in the same chromosome region share common evolutionary events more often than other genes (e.g. a segmental duplication of this region). Their evolution may also be related if they are involved in the same protein complex or biological process. Identifying coevolving genes can thus shed light on ancestral genome structures and functional gene interactions.
Results
We devise a simple, fast and accurate probability method based on species treegene tree reconciliations to detect when two gene families have coevolved. Our method observes the number and location of predicted macroevolutionary events, and estimates the probability of having the observed number of common events by chance.
Conclusions
Simulation studies confirm that our method effectively identifies coevolving families. This opens numerous perspectives on genomescale analysis where this method could be used to pinpoint coevolving gene families and thus help to unravel ancestral genome arrangements or undocumented gene interactions.
Background
Species from the same ecosystem may share common environmental factors (e.g. related to the local climate or to the arrival of new species in the ecosystem) or be interdependent, and their evolution can be related. In the vast majority of cases, the footprint of this dependence is minimal, but in some cases, such as predatorprey, hostparasite or symbiotic relationships, species influence each other so much that their coevolution can be detected [13]. Similarly, nucleotides and amino acids that are located close to one another on the genome share common local factors (e.g. specific nucleotide composition bias or underlying mutation rates due to the functional importance of the locus) and influence each other (e.g. because they are in the same codon, part of the same active site of a protein or because one is part of a transcription factor controlling the transcription level of the other).
The problem of detecting coevolution at the amino acid level has been extensively studied recently ([4, 5]; among others). However, at a broader level, neighbouring genes can also coevolve, sharing common evolutionary events such as segmental duplications [6] and local evolutionary factors such as the proximity of recombination hotspots or centromeres [7]. Protein interactions, e.g. being part of the same protein complex or biological pathway, can also induce coevolution at the gene level. Relatively little work has been done on detecting coevolution at the gene level [812].
To detect gene coevolution, one has to observe it in a significant number of species. As more and more full genomes/transcriptomes are sequenced, more raw data needed to detect coevolving genes becomes available. Being able to accurately detect coevolving genes would, among other things, help to (a) pinpoint possible functional interdependence, allowing us to annotate genomes from nonmodel species; (b) infer ancestral proximity among genes, allowing us to reconstruct ancestral genome arrangements [11]; or (c) cluster genes to reconstruct the Tree of Life in a divideandconquer framework [13, 14].
In [12], Cohen et al. proposed a probabilistic method to detect coevolutionary interactions from phylogenetic profiles, using gain and loss events. They used their method to study a group of 4593 prokaryotic gene families and construct a coevolution network. This yielded several clusters of genes which corresponded to identifiable functional pathways.
In this paper, we propose a novel probabilistic method to detect coevolution. Our method differs from that of [12] in that it is based on species treegene tree reconciliations. Reconciliation methods construct a mapping between a gene tree and a species tree to explain their incongruence by macroevolutionary events such as speciations, gene duplications, horizontal gene transfers etc. Several reconciliation methods have recently been developed following parsimonious or probabilistic paradigms (see [15] for a review). By using reconciliations, we are able to distinguish between different types of events and take into account uncertainties on such events [16, 17].
Our method has advantages over that of [12] in that (a) it can measure coevolution between genes with small or different numbers of events; (b) it can take into account several possible evolutionary scenarios for each gene, reflecting inference uncertainties; and (c) it uses a theoretical modelbased framework to compute pvalues for the coevolution score, rather than bootstrapped simulations as done in [12]. Simulations show that our method is effective in detecting coevolution between genes, even when it is relatively weak. It is also timeefficient, which allows us to conduct genome scale analysis to search for undocumented coevolution among thousands of gene families.
Preliminaries
Let T = (V(T), E(T)) be a (rooted) tree with labelled leaf vertices. We denote the leaves of T by L(T) and the (multi)set of all labels of those leaves by $\mathcal{L}\left(T\right)$. Given a vertex x ∈ V(T), we denote by x_{ p } its parent and by y ≤ x the fact that a vertex y is a descendant of x.
We define a gene tree G as a tree where each leaf represents an extant gene. Likewise, we define a species tree S as a tree in which each leaf represents a distinct extant species. The labels of the leaves of S are unique since they are the identifiers of these species. In gene trees, internal vertices may represent various evolutionary events (e.g. speciation, duplication), while in the species tree they all represent speciation events. In this paper, we suppose that gene and species trees are rooted and binary. Finally, we assume that the genes of G come from the genomes of species present in S, in particular that each label of $\mathcal{L}\left(G\right)$ appears in $\mathcal{L}\left(S\right)$ (denoted by $\mathcal{L}\left(G\right)\u2291\mathcal{L}\left(S\right)$).
A species tree S is said to be dated if it is associated to a function θ_{ S } which represents the time separating a vertex from the current time, i.e. ${\theta}_{S}:V\left(S\right)\to {\mathbb{R}}^{+}$ such that if y ≤ x then θ_{ S }(y) ≤ θ_{ S }(x) and if x ∈ L(S) then θ_{ S }(x) = 0. Using a subdivision of S rather than S itself when computing reconciliations has been proven to ensure timeconsistency of gene transfers in polynomial time [18]. The subdivision S^{′} of S together with an associated time function ${\theta}_{{S}^{\prime}}$ is constructed as follows: firstly, for each node x ∈ V(S) ∖ L(S) and each edge (y_{ p }, y) ∈ E(S) s.t. ${\theta}_{S}\left({y}_{p}\right)>{\theta}_{S}\left(x\right)>{\theta}_{S}\left(y\right)$, an artificial node w is inserted along the edge (y_{ p }, y), with ${\theta}_{{S}^{\prime}}\left(w\right)={\theta}_{S}\left(x\right)$; secondly, for nodes x ∈ V(S^{′}) corresponding to nodes already present in S, we set ${\theta}_{{S}^{\prime}}\left(x\right)={\theta}_{S}\left(x\right)$.
In this paper, we use the combinatorial reconciliation model of Doyon et al.[18], called the $\mathbb{DTL}$ model. We refer the reader to this paper for a formal definition of reconciliations. This model considers (as possible macroevents that shape the genome) speciations, duplications, transfers and losses of genes. For algorithmic reasons losses are never considered alone, so the atomic events of this model are: a speciation ( ), a duplication ( ), a transfer ( ), a transfer followed immediately by the loss of the nontransferred child ($\mathbb{T}\mathbb{L}$), a speciation followed by the loss of one of the two resulting children ($\mathbb{S}\mathbb{L}$), a no event ( ) that only reflects the fact that a gene lineage has crossed a time boundary, and a contemporary event ( ) that associates an extant gene to its corresponding species.
The method of [18] calculates the most parsimonious reconciliation under this model. However, there often exist several most parsimonious reconciliations. Those reconciliations constitute what we call a reconciliation space, which can be efficiently stored in the reconciliation graph introduced by Scornavacca et al.[16].
Methods
In this section we present our new methodology to detect whether or not two gene familes have coevolved. We take as input two gene trees G_{1} and G_{2} and a dated tree S such that $\mathcal{L}\left({G}_{1}\right)\u2291\mathcal{L}\left(S\right)$ and $\mathcal{L}\left({G}_{2}\right)\u2291\mathcal{L}\left(S\right)$.
 1.
We reconcile each of the two gene trees to ${S}^{\prime}$ (the subdivision of S) to produce two corresponding reconciliation spaces. Event sets are then extracted from these two spaces. Details are given in the “Computing the weighted event sets” section.
 2.
We calculate a coevolution score which quantifies the similarity between the two event sets. Details are given in the “Computing the coevolution score” section.
 3.
We calculate the pvalue of the calculated score under a model of independent evolution. If this pvalue is less than an appropriate threshold (reflecting the acceptable error rate for false positive coevolution detection) we consider that G _{1} and G _{2} coevolved. Details are given in the “Computing the pvalue” section.
Computing the weighted event sets
We use the method of [16] to reconcile each of the two gene trees to the subdivided species tree, using equal costs for , and events. This yields two reconciliation spaces RC_{1} and RC_{2} which contain all of the most parsimonious reconciliations between G_{1} (respectively G_{2}) and S. By taking the multiple reconciliations of RC_{1} and RC_{2} into account, we can explore a broad set of possible events, assess their reliability and remove the danger of artifacts arising in a single reconciliation.
Each reconciliation, according to the $\mathbb{D}\mathbb{T}\mathbb{L}$ model, yields a set of events with types from $\{\mathbb{S},\mathbb{D},\mathbb{T},\mathbb{T}\mathbb{L},\mathbb{S}\mathbb{L},\varnothing ,C\}$. However, and events are determined by the species tree, and events are artifacts due to the use of subdivision. Therefore, coincident events of these types are not an indication of coevolution, and we discard them. Likewise, we consider $\mathbb{S}\mathbb{L}$ events only as events. Furthermore, $\mathbb{T}\mathbb{L}$ events are considered as two separate and events.
We are now left with only , and events which we extract from the reconciliation spaces. These events are characterised by their type and their position in the considered gene and species trees. Here, we “undo” the subdivision and consider the position of the event in the original species tree S rather than the subdivided tree ${S}^{\prime}$.
which contains the weights of all events of type on branch b.
Since the frequency of an event over most parsimonious reconciliations has been shown to be a good indicator of its reliability [17], we use ${w}_{1}{(b,u)}_{\mathbb{E}}$ as an estimate of the probability that this event has really occurred in G_{1}. This provides us with a set of possible events together with their probabilities according to G_{1}. Another set is obtained from RC_{2} in a similar way.
Note that these weighted event sets can be obtained from any reconciliation method, for example by taking into account the set of Nearoptimal Parsimonious Reconciliations (NPRs, see [17]), rather than focusing only on most parsimonious reconciliations. Having a set of reconciliations is preferable, since it reflects the inherent uncertainty of reconciliation inference and event prediction. It also allows us to have probability values associated to each event, whereas a single reconciliation only has the presence or absence of events. If only given a single reconciliation, one can also obtain a set of associated (sub)optimal reconciliations, e.g. reconciliations that are reachable by a small number of the operators described in (Chan, Ranwez, Scornavacca: Exploring the space of gene/species reconciliations with transfers. Submitted to J Math Biol).
In fact, we use reconciliations only as a tool to produce the weighted event sets, which are the input to the remainder of the method. In theory, any method which produces a weighted set of genetic events (even if they are not $\mathbb{D}\mathbb{T}\mathbb{L}$ events) can replace this step. We use reconciliations because they provide a straightforward way to calculate the event sets, and there are already efficient algorithms for computing the reconciliation spaces.
Computing the coevolution score
Events of the same type which occur at approximately the same time in both G_{1} and G_{2} support a hypothesis of coevolution. Therefore, we calculate a statistic which measures the amount of coevolution based on the number of such events which are inferred from the reconciliations.
Given two reconciliations — one for G_{1} and one for G_{2} — we could define the coevolution score to be the number of $\mathbb{D},\mathbb{T}$ or events which occur in both reconciliations on the same branch.
 1.
We consider the weight associated to each event ${w}_{1}{(b,u)}_{\mathbb{E}}$ to be the probability that this event has occurred in G _{1}. We make the (strong) assumption that any such event is independent from any other event represented by ${w}_{i}{({b}^{\prime},{u}^{\prime})}_{{\mathbb{E}}^{\prime}}$ for i = 1, 2.
 2.For all branches b ∈ E(S) and element types , we calculate the probability of having 0, 1, …, n events of type on b, where $n=\left{\mathcal{W}}_{1}{\left(b\right)}_{\mathbb{E}}\right$. This is done via recursion as follows: suppose ${\mathcal{W}}_{1}{\left(b\right)}_{\mathbb{E}}=\{{p}_{1},\dots ,{p}_{n}\}$. Let X _{ i } be a variable representing the number of actual events from the first i possible events represented in this set. Then for i = 1, …, n and x = 0, …, i, we have$P({X}_{i}=x)={p}_{i}P({X}_{i1}=x1)+(1{p}_{i})P({X}_{i1}=x),$
 3.
We do the same for G _{2}, using the notations Y and m instead of X and n. The variables X _{ n } and Y _{ m } represent the total number of actual events of type on this branch.
 4.For all branches b of S and element types , we compute the expected number of events in common:$\begin{array}{l}E\left(\text{number of events in common}\right)\\ \phantom{\rule{1em}{0ex}}=\sum _{x=0}^{n}\sum _{y=0}^{m}min(x,y)P({X}_{n}=x)P({Y}_{m}=y).\end{array}$
We define the coevolution score between G_{1} and G_{2} given S as the sum of this value over all branches of S and event types.
Example probability calculation 1
G _{1}  0  1  2  3 

X _{1}  0  1  
X _{2}  0  0.5  0.5  
X _{3}  0  0.25  0.5  0.25 
Example probability calculation 2
G _{2}  0  1  2 

Y _{1}  0.4  0.6  
Y _{2}  0.32  0.56  0.12 
Computing the pvalue
The coevolution score measures the dependence between two gene trees given a species tree. However, its distribution is highly dependent on the number of events in each reconciliation space. In order to assess the significance of the score, we compute the pvalue associated to it.
To do so, we count the average number of events in each event set, which we denote (rounded up) by N_{1} and N_{2}. For each branch b ∈ E(S) and event type , we call the combination $(b,\mathbb{E})$ a bin, and denote by B the (arbitrarily) ordered vector containing all possible bins, over all branches b of the tree S and the 3 element types of . We denote the lengths (representing duration) of the respective branches in S by l_{1}, …, l_{ N }, where N = 3E(S) is the number of bins. In this sequence, each branch length will occur 3 times, once for each event type.
We compute the pvalue under a model that assumes that the genes do not coevolve and all , and events are distributed at random among the elements of B, with probabilities proportional to the branch lengths. Using a theoretical model allows us to efficiently calculate pvalues without simulations which rely on bootstrapped data (as was done in [12]). This increases the reliability of the calculations and mitigates the influence of the independence assumption made when computing the coevolution score (previous section, step 1 of the procedure).
Definition 1.
We define f(x;n_{1}, n_{2}, n) to be the probability that, if n_{1} and n_{2} events are randomly placed on the first n bins of B, there will be at least x events in common between the two event sets.
The variable i in the outside sum denotes the number of events in common between the two event sets in bin n. The first term considers the case where there are exactly i events in this bin in both sets. The second term accounts for the case where the first set has j > i events in this bin, but the second set only has i such events — the number of events in common is still i. The third term considers the mirrored version of the second term.
To calculate f(X;N_{1}, N_{2}, N), we calculate f(x;n_{1}, n_{2}, n) for all x ≤ X, n_{1} ≤ N_{1}, n_{2} ≤ N_{2}, n ≤ N, in order of increasing n. We can do this because (1) expresses f(x;n_{1}, n_{2}, n) in terms of f values where the fourth argument is n  1 and the other arguments are not increased. The lower f(X;N_{1}, N_{2}, N) is, the stronger the evidence against the hypothesis that the genes did not coevolve. To test the coevolution hypothesis, we compare this number to a predefined threshold level, in general 0.05.
Note that the function f itself depends only on the species tree; only its arguments depend on the gene trees and coevolution score. Because of this, we only have to perform the recursion once for every species tree, with the arguments set to the maximal values encountered in the set of genes. This allows us to quickly compute the values of the function for many genes which belong to the same species (which occurs, for example, in our simulations), and so process whole genome analysis to scan for undocumented gene family coevolution.
Results and discussion
In this section, we first describe the simulation protocol used to mimic gene family coevolution along a species tree. We then provide and discuss the results obtained by our method on this dataset, which confirm its ability to detect when two gene families coevolve.
Gene tree simulation
We start with a dated species tree S. Every branch of S has an associated activity a — representing the overall rate at which / / events occur on this branch — and specific rates for each individual event type ${r}_{\mathbb{D}},{r}_{\mathbb{T}},{r}_{\mathbb{L}}$, with $a={r}_{\mathbb{D}}+{r}_{\mathbb{T}}+{r}_{\mathbb{L}}$. We simulate two gene trees simultaneously, with a parameter c ∈ [0, 1] (which we call the coevolution parameter) representing the dependence between the two genes. Informally, an event in one gene tree has a probability c of also occurring in the other gene tree. For example, if c = 1 then the two trees must be identical, whereas if c = 0 they are completely independent.
To simulate the gene trees, we use a modified birthanddeath process which explicitly controls the coevolution between the two genes. At the beginning of the process, the two genes are located at the root of S and paired (identified) to each other. At any time, the time t_{ next } of the next / / event in every existing unpaired gene is calculated by simulating an exponential variable with parameter equal to the activity of the branch (x, y) containing that gene. For gene pairs, this activity must be multiplied by a factor of $\frac{2}{1+c}$ for reasons that will be explained shortly. Then, if t_{ next } ≤ θ_{ S }(y), the next event is determined to be a event if y is a leaf, and an event otherwise. If t_{ next } > θ_{ S }(y), the next event is a / / event and we rely on the relative rates ${r}_{\mathbb{D}},{r}_{\mathbb{T}},{r}_{\mathbb{L}}$to determine its type. If this event affects a gene pair, then:

If it is an , both genes in the pair must speciate. The left (respectively right) child of one resulting gene is then paired to the left (resp. right) child of the other.

If it is a , the event will occur in one gene of the pair with probability 1, and in the other with probability c. If it occurs in both genes, the children are paired to each other as in the case. If it occurs in only one gene, one of the resulting children is paired to the other gene (it does not matter which child).

If it is a , we treat it the same as for a event, with the added conditions that if it occurs in both genes, the transfer targets must be the same, and if it occurs in only one gene, the child which remains in the originating branch is paired to the other gene.

If it is an , the event will occur in one gene of the pair with probability 1, and in the other with probability c. If it occurs in only one gene, the other gene is now unpaired.
It is now clear why the activity of a gene pair above is multiplied by $\frac{2}{1+c}$: each / / event in a pair results in 1 + c actual gene events on average between the two trees. To achieve the correct marginal activity in each gene tree, we must multiply by the correcting factor.
Simulation results
We ran simulations using a phylogeny of 37 proteobacteria over a period of 500 million years as a species tree. We generated duplication, transfer and loss rates for each simulated gene independently, using the same scheme as [19]: the loss rate was randomly chosen in the interval [0.001, 0.0018], where the units are events per gene per million years; the ratio between the “birth” rate (sum of the duplication and transfer rates) and the loss rate was randomly chosen in the interval [0.5,1.1]; finally the proportion of the duplication rate to the birth rate was randomly chosen in the interval [0.7,1]. Both the species tree and the event rates were chosen in accordance with real dataset observations [20].
We observe that the pvalue 1 is overrepresented in all plots. This arises from the granularity of the simulations. More specifically, the pvalue does not come from a continuous distribution, but from a variety of discrete distributions depending on N_{1} and N_{2}, each with a moderate number of possible values. 1 is always one of these values (for when X = 0, i.e. there are no events in common), and so it is overrepresented. This effect is more noticeable as c becomes smaller, because the likelihood of having no event in common grows larger.
It is apparent from Figure 2 that the pvalue statistic is effective in distinguishing between coevolving gene families and independent gene families. Even with quite low values of c such as 0.2, the distribution of the pvalues is noticeably skewed towards 0. At higher levels of c, almost all the pvalues are very close to 0.
If our underlying model is correct, then the case c = 0 in Figure 2 should have a uniform distribution. Even if we ignore the pvalues of 1, our sample distribution is clearly not uniform (a χ^{2} goodnessoffit test to a uniform distribution rejects this hypothesis with a pvalue of less than 10^{15}). This is almost certainly due to the fact that our model assumptions are not an exact match for reality (or, indeed, our simulation protocol). However, the distribution is close enough to uniform that our assumptions appear to be reasonable. In fact the false positive rate for a threshold of 0.05 is only 0.024, less than expected under the underlying model.
Further simulations (which we do not show the results of here) indicate that varying the event costs used in the reconciliation algorithm does not significantly impact these results.
Comparison with the method of Cohen et al.
For a complete assessment of the effectiveness of our coevolution detection algorithm, we compare it to the method of Cohen et al.[12] (henceforth referred to as Cohen’s method) on our simulated data.
We must stress that the two methods accept different input formats; while our algorithm takes gene trees as input, Cohen’s method only uses phyletic patterns of gene presence/absence in extant species, which can be extracted from the gene trees but do not contain all of their information. As such, we should expect our method to outperform Cohen’s method as it uses more information. On the other hand, the fact that our method requires more information as input is not a huge drawback, as full gene tree information is becoming more and more available in recent times.
We ran Cohen’s method on smaller test sets (1000 gene tree pairs) of simulated data for the coevolution parameter values c = 0, 0.2, 0.5, 0.7; the smaller size was for efficiency reasons and is not expected to skew the results. Firstly, because Cohen’s method only compares genes with “exchangeability” (number of inferred gain/loss events) greater than some minimum value, only a small proportion (less than 15%) of the gene families were actually compared. Our method, which can compare any two gene trees, is clearly superior in this respect.
Comparison with the method of Cohen et al.[12]
c  Number of pairs compared by Cohen’s method (out of 1000)  Proportion of pairs withpvalue < 0.05 (Cohen’s method)  Proportion of pairs withpvalue < 0.05 (our method) 

0  68  0  0.024 
0.2  80  0.08  0.247 
0.5  133  0.56  0.762 
0.7  144  0.92  0.930 
Conclusion
In this paper, we have devised an algorithm to detect and measure the strength of coevolution between two gene families. It takes two gene trees as input, and uses their reconciliations to a common species tree to assess the coevolution of the gene families. Simulation studies, and a comparison with the method of Cohen et al.[12], show that this test is an effective way of detecting coevolution.
The detection of strong coevolution among gene families can signal either a proximity or a functional relationship between the families. If working on a fully sequenced genome, the identification of coevolution signals between distant genes could pinpoint ancestral genome rearrangements and/or strong functional links between those genes. If the genome is not fully sequenced, further study may be required to investigate the reason for coevolution and to distinguish between proximity and functional relationships.
Further work includes the design of a clustering method based on coevolution scores to provide biologists with clusters of coevolving gene families rather than just pairwise coevolution information. Another possible avenue for exploration includes extending the current method to include 3 or more gene families. We also plan, in collaboration with experts in bacterial evolution, to apply this method to the bacterial gene trees available in the HOGENOM database [21] to detect existing coevolution among distant genes and to use this information to provide functional insights on unannotated gene families.
Declarations
Acknowledgements
We thank Eric Tannier and Vincent Daubin for fruitful discussions, and Ofir Cohen for his help with the method in [12]. This work was partially funded by the French Agence Nationale de la Recherche Investissements d’avenir / Bioinformatique (ANR10BINF0102, Ancestrome). This publication is contribution no. 2013162 of the Institut des Sciences de l’Evolution de Montpellier (ISEM, UMR 5554).
Authors’ Affiliations
References
 Huelsenbeck JP, Rannala B, Yang Z: Statistical tests of hostparasite cospeciation. Evolution. 1997, 51 (2): http://dx.doi.org/10.2307/2411113,Google Scholar
 Legendre P, Desdevises Y, Bazin E: A statistical test for hostparasite coevolution. Syst Biol. 2002, 51 (2): 217+10.1080/10635150252899734. http://dx.doi.org/10.1080/10635150252899734,View ArticlePubMedGoogle Scholar
 Tangled Trees: Phylogeny, Cospeciation, and Coevolution,. Edited by: Page RDM. 2002, Chicago: University Of Chicago Press, http://www.amazon.com/exec/obidos/redirect?tag=citeulike0720&path=ASIN/0226644677,Google Scholar
 Dutheil J, Pupko T, JeanMarie A, Galtier N: A modelbased approach for detecting coevolving positions in a molecule. Mol Biol Evol. 2005, 22 (9): 19191928. 10.1093/molbev/msi183. http://mbe.oxfordjournals.org/content/22/9/1919.abstract,View ArticlePubMedGoogle Scholar
 Ashkenazy H, Kliger Y: Reducing phylogenetic bias in correlated mutation analysis. Protein Eng Des Sel. 2010, 23 (5): 321326. 10.1093/protein/gzp078. http://peds.oxfordjournals.org/content/23/5/321.abstract,View ArticlePubMedGoogle Scholar
 Koszul R, Caburet S, Dujon B, Fischer G: Eucaryotic genome evolution through the spontaneous duplication of large chromosomal segments. Embo J. 2004, 23: 234243. 10.1038/sj.emboj.7600024. [Koszul et al 2004 eucaryotic genome evolution through duplic large segments.pdf 02614189 Journal Article]PubMed CentralView ArticlePubMedGoogle Scholar
 Zhang L, Gaut BS: Does recombination shape the distribution and evolution of tandemly arrayed genes (TAGs) in the arabidopsis thaliana genome?. Genome Res. 2003, 13 (12): 25332540. 10.1101/gr.1318503. http://www.genome.org/cgi/content/abstract/13/12/2533,PubMed CentralView ArticlePubMedGoogle Scholar
 Barker D, Pagel M: Predicting functional gene links from phylogeneticstatistical analyses of whole genomes. PLoS Comput Biol. 2005, 1: e310.1371/journal.pcbi.0010003. http://dx.doi.org/10.1371%2Fjournal.pcbi.0010003,PubMed CentralView ArticlePubMedGoogle Scholar
 Ané C, Larget B, Baum DA, Smith SD, Rokas A: Bayesian estimation of concordance among gene trees. Mol Biol Evol. 2007, 24 (2): 412426. http://mbe.oxfordjournals.org/content/24/2/412.abstract,View ArticlePubMedGoogle Scholar
 Cordero OX, Snel B, Hogeweg P: Coevolution of gene families in prokaryotes. Genome Res. 2008, 18 (3): 462468. 10.1101/gr.6815508. http://genome.cshlp.org/content/18/3/462.abstract,PubMed CentralView ArticlePubMedGoogle Scholar
 Birin H, Tuller T: Efficient algorithms for reconstructing gene content by coevolution. BMC Bioinformatics. 2011, 12 (Suppl 9): S1210.1186/1471210512S9S12. http://www.biomedcentral.com/14712105/12/S9/S12,PubMed CentralView ArticlePubMedGoogle Scholar
 Cohen O, Ashkenazy H, Burstein D, Pupko T: Uncovering the coevolutionary network among prokaryotic genes. Bioinformatics. 2012, 28 (18): i389i394. 10.1093/bioinformatics/bts396. http://bioinformatics.oxfordjournals.org/content/28/18/i389.abstract,PubMed CentralView ArticlePubMedGoogle Scholar
 Gordon AG: Consensus supertrees: the synthesis of rooted trees containing overlapping sets of labelled leaves. J Classif. 1986, 3: 335335Google Scholar
 BinindaEmonds ORP: Supertree construction in the genomic age. Methods Enzymol. 2005, 395: 745757.View ArticlePubMedGoogle Scholar
 Doyon JP, Ranwez V, Daubin V, Berry V: Models, algorithms and programs for phylogeny reconciliation. Brief Bioinform. 2011, 12 (5): 392400. 10.1093/bib/bbr045.View ArticlePubMedGoogle Scholar
 Scornavacca C, Paprotny W, Berry V, Ranwez V: Representing a set of reconciliations in a compact way. J Bioinform Comput Biol. 2013, 11:Google Scholar
 Nguyen TH Ranwez, Berry V, Scornavacca C: Support measures to estimate the reliability of evolutionary events predicted by reconciliation methods. PLoS ONE. 2013, 8 (10): e7366710.1371/journal.pone.0073667. doi:10.1371/journal.pone.0073667PubMed CentralView ArticlePubMedGoogle Scholar
 Doyon JP, Scornavacca C, Gorbunov KY, Szöllősi GJ, Ranwez V, Berry V: An efficient algorithm for gene/species trees parsimonious reconciliation with losses, duplications and transfers. Proceedings of the 2010 International Conference on Comparative Genomics, RECOMBCG’10. 2011, Berlin, Heidelberg: SpringerVerlag, 93108. http://dl.acm.org/citation.cfm?id=1927857.1927866,Google Scholar
 Nguyen TH, Ranwez V, Pointet S, Chifolleau AMA, Doyon JP, Berry V: Reconciliation and local gene tree rearrangement can be of mutual profit. Algorithms Mol Biol. 2013, 8: 1210.1186/17487188812.PubMed CentralView ArticlePubMedGoogle Scholar
 David LA, Alm EJ: Rapid evolutionary innovation during an Archaean genetic expansion. Nature. 2010, 469 (7328): 9396.View ArticlePubMedGoogle Scholar
 Penel S, Arigon AM, Dufayard JF, Sertier AS, Daubin V, Duret L, Gouy M, Perriere G: Databases of homologous gene families for comparative genomics. BMC Bioinformatics. 2009, 10 (Suppl 6): S310.1186/1471210510S6S3.PubMed CentralView ArticlePubMedGoogle Scholar
Copyright
This article is published under license to 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.