## Figures

## Abstract

Hierarchical orthologous groups are defined as sets of genes that have descended from a single common ancestor within a taxonomic range of interest. Identifying such groups is useful in a wide range of contexts, including inference of gene function, study of gene evolution dynamics and comparative genomics. Hierarchical orthologous groups can be derived from reconciled gene/species trees but, this being a computationally costly procedure, many phylogenomic databases work on the basis of pairwise gene comparisons instead (“graph-based” approach). To our knowledge, there is only one published algorithm for graph-based hierarchical group inference, but both its theoretical justification and performance in practice are as of yet largely uncharacterised. We establish a formal correspondence between the orthology graph and hierarchical orthologous groups. Based on that, we devise GETHOGs (“Graph-based Efficient Technique for Hierarchical Orthologous Groups”), a novel algorithm to infer hierarchical groups directly from the orthology graph, thus without needing gene tree inference nor gene/species tree reconciliation. GETHOGs is shown to correctly reconstruct hierarchical orthologous groups when applied to perfect input, and several extensions with stringency parameters are provided to deal with imperfect input data. We demonstrate its competitiveness using both simulated and empirical data. GETHOGs is implemented as a part of the freely-available OMA standalone package (http://omabrowser.org/standalone). Furthermore, hierarchical groups inferred by GETHOGs (“OMA HOGs”) on >1,000 genomes can be interactively queried via the OMA browser (http://omabrowser.org).

**Citation: **Altenhoff AM, Gil M, Gonnet GH, Dessimoz C (2013) Inferring Hierarchical Orthologous Groups from Orthologous Gene Pairs. PLoS ONE 8(1):
e53786.
https://doi.org/10.1371/journal.pone.0053786

**Editor: **Olivier Lespinet,
Université Paris-Sud, France

**Received: **August 22, 2012; **Accepted: **December 5, 2012; **Published: ** January 14, 2013

**Copyright: ** © 2013 Altenhoff et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

**Funding: **This work was supported by an ETH Independent Investigators’ Research Award to CD and GHG, and an SNSF advanced researcher fellowship to CD (#136461). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

**Competing interests: ** The authors have declared that no competing interests exist.

## Introduction

Homologous biological sequences–sequences related through common ancestry–can be further classified according to the type of evolutionary event that initiated their divergence from one another. Notably, pairs of genes that descended from their last common ancestor through a speciation are referred to as orthologs, while genes that have diverged from a duplication event are referred to as paralogs [1]. This distinction is useful in a broad range of contexts, such as genome annotation, comparative genomics, and phylogenetic analyses. Accordingly, numerous methods and associated databases have been developed to infer orthology and paralogy (reviewed in [2], [3]).

Orthology between pairs of genes can be quite reliably inferred using various algorithms, such as bidirectional best hit [4], reciprocal smallest distance [5], Inparanoid [6], or OMA pairwise [7] (see [8] for in-depth description and evaluation). Yet, many analyses require relations over more than two genes at a time. But because in general orthology and paralogy are non-transitive relations (i.e. x being orthologous to y and y being orthologous to z does not imply x being orthologous to z), the generalisation of these concepts to sets of genes is not straightforward. As a consequence, several definitions of orthologous groups have been proposed, with considerable differences in terms of evolutionary relations implied [9].

For instance, OrthoMCL identifies groups of orthologs and “close” paralogs using Markov clustering, a procedure to identify sets of genes with high pairwise alignment scores [10]. A more stringent grouping strategy lies in identifying cliques of orthologs, but this comes at the cost of lower coverage in terms of all orthologous relations [7]. Also worth mentioning are criteria which are not directly aiming for orthology, such as groups with a given minimum percentage of sequence identity (e.g. [11], [12]) or minimum percentage of sequence length coverage (e.g. [12]). However, these non-evolutionary criteria can yield groupings which are at odds with the central notion of evolution and function changes along trees.

One particularly useful gene grouping strategy, sometimes referred to as *hierarchical orthologous groups*, entails grouping genes that have descended from a single common ancestral gene in the last common ancestral species of a given taxonomic range [Figure 1]. This definition has several interesting implications: (i) defining groups in terms of specific taxonomic ranges enables users to fine-tune their analyses to different contexts of investigation–for instance, studying “ubiquitous” genes among all species, but studying lactation genes in terms of the last mammalian common ancestor only; (ii) hierarchical groups have a straightforward interpretation in terms of gene trees: they are clades on these trees; (iii) collectively, hierarchical groups defined with respect to every ancestral species capture all orthologous and paralogous relations [9]. Hierarchical groups are at the heart of the orthology databases EggNOG [13] and OrthoDB [14]. Since 2011, we also provide this type of grouping in the OMA database [15].

In this example, the hierarchical groups for the taxonomic range are drawn in orange. By definition, these groups correspond to the sets of leaves attached to the speciation nodes of the gene tree coloured in orange.

Hierarchical groups can be trivially derived from reconciled gene/species trees, such as those obtained by LOFT [16], Ensembl Compara [17], Synergy [18], or PhylomeDB [19]. However, these tree-based approaches are computationally expensive, and indeed, most large phylogenomic databases are “graph-based”, i.e. they infer orthology based on pairwise gene comparison [3]. But most well-established graph-based methods do not attempt to reconstruct hierarchical groups: OrthoMCL groups are a trade-off between orthology and paralogy [20], Inparanoid works for pairs of genomes only [21] and RoundUp computes pairs of orthologs only, i.e. there is no further grouping [22]. To our knowledge, the only graph-based algorithm for hierarchical group inference published to date is COCO-CL [23], which despite its pioneering character is a somewhat ad-hoc approach. Indeed COCO-CL can be shown to return suboptimal results on relatively simple problems, even with perfect input data (example provided in Materials S1).

In this article, we present GETHOGs, which stands for “Graph-based Efficient Technique for Hierarchical Orthologous Groups”. The algorithm is based on correspondences between the orthology graph and the underlying gene phylogeny, correspondences that we prove in two new lemmas. We present an efficient implementation of the algorithm as part of the OMA standalone package. We demonstrate that the resulting algorithm outperforms COCO-CL on simulated and real data. We also show that GETHOGs outperforms the tree reconciliation method LOFT. Lastly, we contrast GETHOGs’s results on real data with predictions of the EggNOG and OrthoDB databases (whose precise algorithms are as yet unpublished).

## Methods

In this section, we first mathematically define hierarchical orthologous groups in terms of gene and species trees, and derive useful notions and properties. We then define the orthology graph, which, crucially, can be inferred without computing gene trees. Next, we describe the correspondence between hierarchical orthologous groups and the orthology graph. The rest of the section details the data and methods used for validating and comparing our new algorithm with existing approaches.

Readers not interested in the technical details can skip this section and proceed directly to the description of GETHOGs (Results Section).

### Labelled Gene Trees, Species Trees, and Hierarchical Orthologous Groups

Let be a forest of rooted gene trees where the internal nodes are labelled either as speciation or duplication nodes. We denote the speciation nodes on these gene trees as . Note that one speciation *event* at the genome level (i.e. on the species tree) corresponds to multiple speciation nodes on the gene trees (one node per ancestral gene in the ancestral species undergoing speciation). The leaves of the trees represent present day genes, so that we use the terms leaf and present day gene interchangeably. Furthermore, we require that each gene belongs to exactly one gene tree in . With , we denote the last common ancestral gene of two present day genes and . For any present day gene , the operator denotes the species that gene belongs to, where is the set of all species covered by . Similarly, we denote by the leaves attached to a speciation node the set of leaves contained in the subtree rooted by the speciation node and by the taxonomic range of , represented by the set of species appearing in the subtree rooted at speciation node (see Figure 1 for illustration of some of these definitions).

By definition, –the set of genes that have descended from the speciation node –constitutes one *hierarchical group* for the taxonomic range . For example, consider an ancestral gene that has duplicated within the vertebrates, but before the mammalian radiation. For the vertebrate taxonomic range, all present day genes that have descended from that ancestral gene are in the same hierarchical group, say , assuming corresponds to the first speciation event among the vertebrates; by contrast, with respect to the mammals, these present day genes are split into two distinct hierarchical groups, say and , where the speciation nodes and correspond to the first speciation event among the mammals.

Let be a chosen *taxonomic range*, such that forms a monophyletic group of size in the species tree. Such a subset of species induces a subset of present day genes . We define to be the minimal set of speciation nodes whose leaves (i.e. hierarchical groups) collectively include most genes in without including genes in species outside . A constructive definition of is given by Algorithm 1 (Table 1). Note that in the absence of gene loss, the nodes are the ancestral genes in the last common ancestor of , which is a more intuitive way of thinking about this set. From these speciation nodes, we derive i) as the forest of subtrees of rooted at the speciation nodes ; and ii) , the set of hierarchical orthologous groups induced by speciation nodes .

#### Proposition 1.

*For all speciation nodes , there is no speciation node with the two following properties: i) is an ancestral node of ; ii) corresponds to a speciation event within the taxonomic range .*

#### Proof.

We prove this proposition by contradiction. Assume the existence of such speciation nodes . Since is an ancestral node of , . Furthermore, since is within the taxonomic range , the additional leaves of belong to species in as well (i.e. ). But then, is not part of the minimum set of speciation nodes whose leaves collectively cover most of without covering genes in species outside , which contradicts our assumption.

#### Proposition 2.

*The correspondences between the speciation nodes in , the trees in , and the hierarchical orthologous groups are all one-to-one.*

#### Proof.

The one-to-one correspondence between and can be established as follows: recall that we require the leaves of the forest to be distinct; thus, each element in is the root of a distinct gene subtree in . Conversely, each tree in has a distinct speciation node as root. Likewise, the one-to-one correspondence between and also follows from the requirement that leaves in be distinct: this guarantees that the hierarchical orthologous groups associated with each speciation node are distinct; furthermore, by definition, each element in is constructed from a distinct speciation node in . Finally, the correspondence between and can be viewed as a composition of the two previous one-to-one correspondences, and is therefore one-to-one itself.

#### Proposition 3.

*If two present day genes belong to distinct hierarchical orthologous groups in , and are not orthologous.*

#### Proof.

Let be the speciation node (ancestral to gene ) corresponding to hierarchical orthologous group , and the speciation node (ancestral to gene ) corresponding to . Since and given Proposition 2, and are distinct. We show by contradiction that and cannot be orthologous. Assume that and are orthologous. Hence, by Fitch’s definition of orthology, and are related through a speciation node , which, since the two genes belong to , corresponds to a speciation event within the taxonomic range . Furthermore, their respective ancestral nodes and are distinct, which means that must be ancestral to and . But Proposition 1 states that there is no such speciation node, which contradicts our assumption.

### The Orthology Graph

We define an *orthology graph* to be a graph over the present day genes as nodes and with edge set , representing pairwise orthology relations between genes as defined by Fitch [1], i.e. they are symmetric, but non-transitive. We further require that every present day gene in be part of at least one orthologous relation, such that has no singleton. As mentioned in the introduction, pairwise orthologs can be inferred using well-established methods, many of which do not require gene tree reconstruction or gene/species tree reconciliation.

Here, we consider two cases: perfect data, where we assume that the pairwise orthologs have been correctly and exhaustively identified, and “real data”, where these have been imperfectly identified, using OMA pairwise (Sect. “Orthology graph inference”; [7] ).

To restrict the orthology graph to a chosen taxonomic range, we denote by the orthology subgraph induced by the vertex subset , again, without singleton genes. Finally, denotes the set of connected components in . A connected component is defined as a maximal subgraph where there exists a path on the graph between every pair of nodes.

### Correspondence between Hierarchical Orthologous Groups and Orthology Graph

Our novel algorithm for hierarchical orthologous group inference will use the following two lemmas. The first lemma establishes a correspondence between hierarchical orthologous groups and the orthology graph (illustrated in Figure 2):

Likewise, it also induces an orthology subgraph with set of connected component (right). Lemma 1 establishes the one-to-one correspondence between and (which we prove by viewing it as composition of the one-to-one correspondences and ).

#### Lemma 1.

*Given a taxonomic range , there is a one-to-one correspondence between the hierarchical orthologous groups and the connected components of the orthology subgraph of the taxonomic range in question .*

#### Proof.

As Proposition 2 asserts, the correspondence between the hierarchical orthologous groups and the speciation nodes is one-to-one. Thus, it suffice to establish a one-to-one correspondence between and . Recall that by definition (and as illustrated in [Figure 1] ), genes in are both the leaves attached to the speciation nodes and the nodes in . This defines a correspondence between the two sets, which we now demonstrate is one-to-one. First, each element is the root of a tree in , and thus has at least two leaves attached to it; in turn, these leaves belong to at least one . Conversely, any has at least one present day gene; in turn, this gene belongs to at least one whose root is by definition in . Next, we show that no is paired with more than one . Let be the tree rooted at . The left and the right subtrees of the speciation node partition the genes in into two sets. By definition, all the genes in one set are orthologous to all the genes in the other set. Therefore, the two sets form a complete bipartite subgraph of and, hence, lie in one connected component. To conclude the proof, we show by contradiction that no is paired with more than one . Assume the existence of a connected component paired with speciation nodes in . As Proposition 2 establishes, this implies that the connected component is paired with orthologous groups in . However, Proposition 3 asserts that all pairs of genes belonging to different such groups are non-orthologous, and thus are not connected by an edge in . But then, there can be no edge between the subsets of genes of that belong to different groups, which contradicts our assumption that is a connected component.

In the second lemma, we prove that on perfect data, members of a hierarchical group have at most two degrees of separation in the orthology graph. Intuitively, this can be seen by the fact that the deepest split in all considered gene (sub)trees is a speciation node, so every gene in one subtree of this split is orthologous to every gene in the other subtree of that split. Hence, regardless of the relationships within these subtrees, it is always possible to go to another gene within the same subtree by first going to any gene in the other subtree and then coming back.

#### Lemma 2.

*Each connected component in has a diameter of at most 2, i.e. every pair of genes within a hierarchical group is separated by at most 2 edges.*

#### Proof.

According to Lemma 1, every maps onto a . The left and the right subtree of the speciation node partition the genes into two sets. By definition, all the genes in one set are orthologous to all the genes in the other. Therefore, contains a complete bipartite graph known to have at most diameter 2.

We will make use of Lemma 2 to motivate and establish the heuristic FractionReachableInTwoSteps parameter to cope with imperfect input data.

### Methods and Data for Validation and Comparison

#### Simulated genomes.

To generate the simulated genomes we used ALF [24], which simulates events at both gene-level (substitution, indels) and genome-level (gene duplication, speciation). For the present work, we simulated 4 independent runs for two different parameter sets: the root genome of each simulation consisted of 200 randomly and independently generated sequences with -distributed lengths. Although 200 genes is much fewer than in most real genomes, the present work pertains to evolutionary relations *within* homologous families, not among them; as such, the number of starting genes can be viewed as the number of replicates we use to obtain result averages. Sequence evolution was simulated with two M3 codon models [25] with default parameters along a species tree of 30 taxa sampled from a birth-death process with birth rate and death rate 0.001. The distance from the root to the leaves was set to 150 PAM. Gene duplication and loss rates both were set to for the first parameter set. For the second parameter set, we set duplication rate to and loss rate to 0.005. In the second parameter set, we additionally allowed temporal rate changes after duplication to model sub- and neo-functionalization of genes as well as gene fusion and fission. The ALF parameter files with all options are provided as supplementary materials (Datasets S1 and S2).

#### Empirical data.

We reanalysed the three gene families from a recent manually curated study by Boeckmann et al. [9]. As input for our algorithm, we used the pairwise orthologs from the OMA May 2010 release as orthology graph and the NCBI taxonomy [26] as the species tree.

#### Orthology graph inference.

To construct the orthology graph, we used pairwise orthologs inferred by the OMA algorithm [7], which has been shown to be competitive in benchmarking studies [8], [9], [27], [28]. In brief, the OMA algorithm first computes all-against-all sequence alignments using full dynamic programming. From these, potential orthologs (“stable pairs”) are selected based on evolutionary distances and considering inference uncertainty. In a verification step, the algorithm identifies pseudo-orthologs arising through differential gene loss [29]. The resulting “verified pairs” are used to construct the orthology graph for the hierarchical clustering method proposed here.

#### Species-tree inference.

With the simulated dataset, we do not assume knowledge of the true species tree. Instead, we estimate it using a least-squares distance approach (*MinSquareTree()* function in Darwin; [30]), using OMA groups as sets of marker genes [7].

#### COCO-CL on COG clusters.

COCO-CL requires initial homologous clusters and refines them into a hierarchy by applying a single linkage clustering algorithm on the induced pairwise distance estimates of the cluster’s multiple alignment. As suggested by the authors [23], we built the initial clusters using the COG algorithm [31]. The COG parameters were chosen according to software documentation, i.e. E-value cutoff = and hit coverage threshold = 0.5. We applied COCO-CL on both the simulated and real datasets. On simulated data, and in order to assess the COCO-CL gene family refinement procedure independently from the COG clustering step, we also used the true simulated homologous gene families as input clusters. To conform to the definition of hierarchical groups, we fixed COCO-CL’s paralogy threshold , i.e. two sub-clusters sharing genes from the same species have to be related by a duplication. For the analysis on simulated data, we varied the bootstrap threshold between 0 and 0.95. For the analysis on empirical data, we set the bootstrap threshold to the default value (0.75).

#### LOFT.

LOFT is a tree-based orthology inference method [16]. It computes Neighbour-Joining gene trees based on pairwise distances using the model [32] followed by an evolutionary event-labelling step of the internal nodes based on a species overlap criterion. Similarly to COCO-CL, LOFT requires initial gene families to work on. Again, we use the inferred COG clusters using the parameters as described above on both simulated and real datasets. On the simulated dataset, as additional control, we repeated the analyses using the true and complete homologous gene families as input.

#### Performance metric.

Following Boeckmann et al. [9], we measured the performance of a method in terms of the precision and recall of *pairwise* orthology or paralogy. Precision and recall are defined as and , where is the number of true positive reported relations, the number of spuriously reported relations and the number of missing predictions. Both precision and recall are bound to the interval , with higher values indicating better performance.

## Results and Discussion

We first present an algorithm which, given a perfect input orthology graph (i.e. all the pairwise orthologs have been correctly and exhaustively identified) and the true (partially or fully resolved) species tree topology, correctly identifies for all taxonomic ranges the corresponding hierarchical orthologous groups. In the second part, we present extensions to cope with imperfect data followed by some remarks about the implementation of the algorithm. We conclude this section by comparing the performance of GETHOGs with existing methods.

### GETHOGs Algorithm

#### Perfect input data.

In order to obtain a hierarchy of nested orthologous groups, our approach requires a rooted, at least partially resolved species tree. Our proposed algorithm computes a hierarchy of orthologous groups by recursively identifying the connected components on the orthology subgraphs induced by the species in the lineages at various taxonomic levels (Algorithm 2, Table 2). As Lemma 1 shows, these connected components directly correspond to hierarchical groups. The corollary to this is that other clustering criteria are suboptimal in at least some cases. For example, the COG triangle-based algorithm is too restrictive when we reach the hierarchical level of two species only. At the same time, it can erroneously merge different groups if they are related through speciation events outside the taxonomic range of interest. More stringent clustering approaches (e.g. MCL with typical parameters) will fail in other cases.

Note that, due to the definition of hierarchical groups, genes belonging to different groups at the same taxonomic range have descended from distinct genes in the last common ancestor. As we have formally established in Proposition 3, such genes are in no circumstance orthologous, and are paralogous if the groups are evolutionarily related (homologous).

The runtime complexity of the GETHOGs algorithm on perfect input data is , where denotes the iterated logarithm function (which grows at a much slower rate than the logarithm function itself). Indeed, algorithm essentially traverses the species tree. In the worst case, the species tree is completely unbalanced. Hence, there are at most levels of recursion. Within each recursion, we need to compute the following elements: first, computing the induced subgraph requires visiting every edge in the orthology graph. This can be done in , because that the number of genes is bound by , were is the number of genes in the largest genome. Second, we need to be able to access the children of the root, which can be done in the worst case, a star tree, in . And third, computing the connected components in a graph can be done in , where denotes the inverse Ackermann function [33]. Hence, the time complexity of each level is dominated by this last step, which multiplied by the of recursion gives the overall time complexity.

#### Imperfect input data.

The two Lemmas described in the “Methods” section are only valid for perfect data. In practice, for all but trivial examples, the input orthology graph can be expected to have missing (false negative) and spurious (false positive) orthology predictions. While missing predictions are typically not a problem–the orthology graph is normally dense enough to provide a path from every group member to every other–additional predictions are more disruptive: false positives result in the erroneous merging of orthologous groups. Hence, using the transitive closure of the pairwise orthology relations would in such situations lead to excessively large clusters. Fortunately, these spuriously merged clusters are often not strongly connected to each other, with only few edges connecting them [Figure 3].

An example orthology graph from the OMA database where two false positive prediction merges two well-defined orthologous groups. At the level of vertebrates, the *NOX* family forms 4 different orthologous groups. Because of two spurious predictions, the *NOX1* and *NOX2* clusters get weakly connected. The minimum cut algorithm will split them, as there are only two edges to cut.

To cope with such errors in the orthology graph, we modify/extend the algorithm GETHOGs (Algorithm 2, Table 2) in the following way: We replace the ConnectedComponents function by DivideGraph (Algorithm 3, Table 3) : this procedure divides the orthology graph using a Minimum-Cut algorithm [34] until all the subgraphs conform sufficiently to the property established in Lemma 2. Minimum-Cut is the well-known computer science problem of cutting a graph into two disjoint components by removing the smallest number of edges (or, in the weighted version, edges with the smallest sum of edge weights). Treating our problem as Minimum-Cut is reasonable in that cutting the graph is needed to undo the effect of spurious edges across groups, while the minimum criterion satisfies the parsimony principle.

As for the termination criterion, it is motivated by the property that with correct input, connected components graphs have diameter of at most 2 (Lemma 2). To approximate the diameter, which is expensive to compute, we estimate the average fraction of nodes which are reachable within two steps of each node (Algorithm 4, Table 4). On perfect data, where the diameter is at most 2, this statistic is necessarily equal to 1. On real data, we however allow for values lower than 1, using the stringency parameter . Based on empirical analyses (see below), we have chosen as the default parameter.

Furthermore, it is also possible to use the weighted version of Minimum-Cut. For this purpose, we augment the orthology graph with edge weights corresponding to pairwise alignment scores, and use these weights to guide the Minimum-Cut algorithm. The rationale is that spurious false positives often have relatively low alignment scores. Hence, the spurious edges erroneously connecting two bona fide groups will have low scores and thus be targeted by the weighted Minimum-Cut procedure. But note that while this heuristic has a theoretical motivation based on our findings on perfect data, we do not claim it to be optimal.

We now give an asymptotic runtime analysis of our algorithm. Giving a tight bound on the runtime analysis on imperfect input data is not easy. We therefore make the assumption that gene duplications and losses are distributed uniformly on the gene trees (thus resulting in a mostly balanced gene family tree).

The time complexity of the DivideGraph algorithm depends on that of MinimumCut and FractionReachableInTwoSteps. FractionReachableInTwoSteps runs in order . Essentially, we have to traverse the graph in breadth-first order from a constant set of starting nodes. The algorithm by Karger and Stein [35] finds a minimum-cut in . Hence, the time complexity of FractionReachableInTwoSteps is dominated by MinimumCut.

The depth of the recursion DivideGraph depends heavily on the structure of the orthology graph. Obviously, it is limited by the number of nodes, but generally, many fewer iterations are necessary. With the assumption that duplications and losses are uniformly distributed on the gene tree, we can expect the graph to be partitioned in proportions of the total size. Then, iterations are required, which leads to a time complexity for DivideGraph of .

The resulting overall time complexity for GETHOGs on imperfect data is therefore of order .

### Implementation

The source of the described algorithm is freely available for non-commercial uses as part of the OMA standalone package on http://omabrowser.org/standalone. The implementation is written in *Darwin*, an interpreted computer language tailored for bioinformatics applications [30]. An important part for our algorithm is a fast implementation of the *Minimum Cut* algorithm. As a new part of *Darwin*, we added a C implementation of the randomised minimum cut algorithm by Karger and Stein [35].

The Karger-Stein algorithm was implemented for weighted graphs. The algorithm is randomised, that is to say, with certain probability (which can be made arbitrarily small) it may not find the minimum cut, but one slightly larger than the minimum. In practice, we could not find cases where it failed for the default parameters, and even if it would fail, this would mostly alter the order in which we find the groups. This randomization allows us to parallelise the procedure for very large graphs [Materials S1].

### Comparison with Existing Methods

We applied our algorithm to both simulated and real data problems, and compared them to a graph-based and a tree-based hierarchical grouping strategies. We generated two artificial datasets by simulation with ALF [24]: one with moderate gene duplication rate, the other with high duplication rate, rate changes after duplication and gene fusion, and fissions (see Methods). For graph-based COCO-CL, following the authors’ protocol, we inferred initial COG clusters and refined them using different bootstrap parameters ranging from to [23]. For tree-based LOFT, also following the authors’ protocol, we inferred one gene tree per COG cluster and inferred duplication nodes by species overlap [16]. For GETHOGs, we used the OMA algorithm [7] to obtain a pairwise orthology input graph. To measure the correctness of the inference in the simulated datasets, we compared the reported induced pairwise orthologous and paralogous relations of the two methods to the true relations obtained from the simulation.

On the dataset with moderate duplication rate, compared to COCO-CL and LOFT, GETHOGs reported considerably more orthologous relations at roughly the same level of precision [Figure 4]. With respect to paralogous relations, GETHOGs strongly outperformed the other methods both in precision and recall. We observed similar trends on the dataset with high duplication rate, except that the precision of GETHOGs orthologs was lower than the precision of COCO-CL and LOFT orthologs. In terms of parameter sensitivity, GETHOGs was little affected by the choice of stringency parameter in the first dataset; in the dataset with higher duplication rate, a precision-recall trade-off became apparent, with high values resulting in moderately higher overall precision and lower overall recall. COCO-CL proved to be more sensitive to parameter choice, with low bootstrap parameter values generally yielding better overall performances.

The two datasets show average rates of 4 independent runs of genome simulations with fixed parameters. The difference between the two datasets are essentially different gene duplication rates (see Method section for details). As a point of reference, we also show the performance of pairwise orthologs inferred in OMA (OMA Pairwise). The colour gradient corresponds to various parameter values for GETHOGs and bootstrap value for COCO-CL.

To analyse the sensitivity to the species phylogeny required by GETHOGs, we ran the algorithm once with the true species tree and once with a species tree inferred from the data (Supplementary Figure 1 in Materials S1). On the first dataset, we observed virtually no difference between GETHOGs with inferred and true species tree, while on the second, more difficult dataset, supplying the true tree led to a modest improvement in precision (Supplementary Figure 1 in Materials S1). This analysis suggests that while GETHOGs can benefit from knowing the true species tree, the method remains competitive when the species tree needs to be inferred.

The surprisingly low recall of LOFT with respect to orthologs and paralogs in the more difficult dataset can be mainly attributed to errors in the gene family inference step, for which LOFT uses the COG algorithm. Indeed, if provided perfect gene family input, the recall for LOFT and COCO-CL increases substantially for both orthologs and paralogs (Supplementary Figure 1 in Materials S1). This suggests that in general the performance of gene/species tree reconciliation methods might strongly depend on the initial family clustering step.

We now turn to the evaluation on empirical biological data. With real data, the true evolutionary relations are mostly unknown. Therefore, we restrict our analysis to a small set of thoroughly studied gene families, which we assume to be free of errors [9]. Again, we compared the predicted relations with the induced relations from these labelled reference gene trees.

This analysis covers three gene families, the “ancestral-type” subfamily of NADPH oxidases (*NOX*1-4), the Popeye domain family (*POP*) and the the eukaryotic V-type ATP synthase beta subunit subfamily (*VATB*). All three families contain at least one lineage specific gene duplication but no horizontal transfer and no major change in their single-domain structure.

In this analysis, we observe that the predictions of GETHOGs largely outperform the ones of OrthoDB and COCO-CL in terms of precision and recall [Figure 5]. Compared with EggNOG and LOFT, the differences are more modest: while EggNOG outperforms our method slightly on the *VATB* gene family, our predictions are considerably better for the *POP* family. As for LOFT, it showed similar performance to GETHOGs for the *POP* and *VATB* families, but did noticeably worse than GETHOGs on *NOX* genes.

[**9**]**.** Predictions for *GETHOGs* and *COCO-CL* are computed using the default parameters (respectively and bootstrap). The points for *EggNOG* and *OrthoDB* are from the original analysis (Reference [9],table 2).

Although these 3 families are not sufficient to draw general conclusions, they nevertheless suggest that the good performance of GETHOGs in simulation extends to real data as well. Furthermore, it should be noted that the absence of description of the EggNOG and OrthoDB algorithms, let alone available implementation, precludes their use on custom genomic data.

We finish this section by discussing the limitations of GETHOGs. Most importantly, the method depends on the quality of the input orthology graph. We have established that GETHOGs returns optimal graphs on perfect input data, but we cannot expect perfect input data on real data. Although we have introduced heuristics to cope with errors in the orthology graph, the performance will deteriorate when the input information is not sufficient to discriminate among multiple evolutionary scenarios. We acknowledge that OMA pairwise, which is known to be relatively conservative [8], [28], might not necessarily provide the best input orthologs for GETHOGs; it might for instance be that GETHOGs works better with a more inclusive method, such as Inparanoid (we plan to investigate this question in a later study).

One potential problem with the input graph might be caused by genes encoding multi-domain proteins. Indeed, if the pairwise orthology detection method used to construct the orthology graph does not ensure that orthology between two genes extend over all (or at least most) domains, the resulting graph might strongly violate GETHOGs working assumptions. Note however that the very concept of orthology among genes with different domain composition (and thus non-homologous parts) is ill-defined, as orthology is a subtype of (and thus presupposes) homology. Because of that, many pairwise orthology inference algorithms, including the OMA algorithm we used for all input in this work, require homologous regions between two genes to extend over most of their sequence lengths [7]. Such requirement is sufficient to ensure that there be no orthology inferred between multi-domain genes/proteins with significantly different domain composition.

The other main limitation of GETHOGs lies in the computational cost of processing huge gene families. The currently biggest orthology graph in the OMA database contains genes and ortholog relations, which is prohibitively expensive for GETHOGs. On very large families, we currently circumvent this problem by starting the GETHOGs recursion at more specific taxonomic levels than the root of all species. Practically, this means that we abstain from resolving the deepest orthology/paralogy relationships in such families. Note however that GETHOGs is able to process most gene families in OMA from the root of the species tree. To give an idea of actual runtimes, computing hierarchical groups on a graph of genes and orthologous relations took about 2 minutes on a single desktop computer; processing another graph with genes and took about 15 minutes.

### Conclusion

We presented GETHOGs, a novel algorithm for reconstructing hierarchical orthologous groups. The approach is based on an orthology graph induced by pairwise orthologous gene relations, and as such requires neither gene tree inference nor gene/species tree reconciliation. The algorithm is motivated by a lemma demonstrating the equivalence of the connected components in the orthology subgraph induced by a taxonomic range and the orthologous groups with respect to the same taxonomic range on perfect data. In order to extend the algorithm to be applicable for real data, we separate weakly connected components by splitting the graph repeatedly at its minimum cut. We stop once the graph is sufficiently densely connected, based on the lemma that the orthology graph should have diameter less than or equal to two.

We applied the algorithm on simulated and real datasets, and compared it to COCO-CL and LOFT, where it finds considerably more orthologs/paralogs at roughly the same precision rate. On real data, we also compared our algorithm to EggNOG and OrthoDB–two databases providing hierarchical orthologous groups–by re-analysing three manually curated gene families from a recent study. Though two the empirical datasets are too small to draw general firm conclusions, the results based on these families indicate that our method is competitive.

Regardless of these promising results, the *raison d*’*être* of GETHOGs lies not so much in resolving once and for all the graph-based hierarchical orthologous group problem as in providing a well-founded and useful starting point to tackle this problem. The theoretical results and implementation provided alongside this study will hopefully foster the development of even better solutions.

## Supporting Information

### Materials S1.

PDF containing supplementary materials and figures.

https://doi.org/10.1371/journal.pone.0053786.s001

(PDF)

### Dataset S1.

Parameter file used to generate the simulated datasets with low duplication rate with ALF [24].

https://doi.org/10.1371/journal.pone.0053786.s002

(DRW)

### Dataset S2.

Parameter file used to generate the simulated datasets with high duplication rate with ALF [24].

https://doi.org/10.1371/journal.pone.0053786.s003

(DRW)

## Acknowledgments

We thank Daniel Dalquen (ETH Zurich), Nick Goldman (EMBL-EBI), Kevin Gori (EMBL-EBI), Matthieu Muffato (EMBL-EBI), Paul Thomas (University of Southern California) and Stefan Zoller (ETH Zurich) for valuable input and stimulating discussions.

## Author Contributions

Conceived and designed the experiments: AA CD. Performed the experiments: AA. Analyzed the data: AA CD. Contributed reagents/materials/analysis tools: AA GG MG CD. Wrote the paper: AA MG GG CD.

## References

- 1. Fitch WM (1970) Distinguishing homologous from analogous proteins. Syst Zool 19: 99–113.
- 2. Kristensen DM, Wolf YI, Mushegian AR, Koonin EV (2011) Computational methods for Gene Orthology inference. Briefings in bioinformatics 12: 379–391.
- 3.
Altenhoff AM, Dessimoz C (2012) Inferring orthology and paralogy. In: Anisimova M, editor, Evolutionary Genomics, Clifton, NJ, USA: Methods in Molecular Biology. 259–279.
- 4. Overbeek R, Fonstein M, D’Souza M, Pusch GD, Maltsev N (1999) The use of gene clusters to infer functional coupling. Proc Natl Acad Sci USA 96: 2896–2901.
- 5. Wall DP, Fraser HB, Hirsh AE (2003) Detecting putative orthologs. Bioinformatics 19: 1710–1711.
- 6. Remm M, Storm C, Sonnhammer E (2001) Automatic clustering of orthologs and in-paralogs from pairwise species comparisons. J Mol Biol 314: 1041–52.
- 7. Roth AC, Gonnet GH, Dessimoz C (2008) The algorithm of OMA for large-scale orthology inference. BMC Bioinformatics 9: 518.
- 8. Altenhoff AM, Dessimoz C (2009) Phylogenetic and functional assessment of orthologs inference projects and methods. PLoS Comput Biol 5: e1000262.
- 9. Boeckmann B, Robinson-Rechavi M, Xenarios I, Dessimoz C (2011) Conceptual framework and pilot study to benchmark phylogenomic databases based on reference gene trees. Brief Bioinform 12: 474–484.
- 10. Li L, Stoeckert CJJ, Roos DS (2003) Orthomcl: identi_cation of ortholog groups for eukaryotic genomes. Genome Res 13: 2178–2189.
- 11. Chen C, Natale DA, Finn RD, Huang H, Zhang J, et al. (2011) Representative proteomes: a stable, scalable and unbiased proteome set for sequence analysis and functional annotation. PLoS ONE 6: e18910.
- 12. Miele V, Penel S, Duret L (2011) Ultra-fast sequence clustering from similarity networks with SiLiX. BMC Bioinformatics 12: 116.
- 13. Powell S, Szklarczyk D, Trachana K, Roth A, Kuhn M, et al. (2012) eggNOG v3.0: orthologous groups covering 1133 organisms at 41 different taxonomic ranges. Nucleic Acids Research 40: D284–9.
- 14. Waterhouse RM, Zdobnov EM, Tegenfeldt F, Li J, Kriventseva EV (2011) OrthoDB: the hierarchical catalog of eukaryotic orthologs in 2011. Nucleic Acids Research 39: D283–8.
- 15. Altenhoff AM, Schneider A, Gonnet GH, Dessimoz C (2011) OMA 2011: orthology inference among 1000 complete genomes. Nucleic Acids Res 39: D289–D294.
- 16. van der Heijden RT, Snel B, van Noort V, Huynen MA (2007) Orthology prediction at scalable resolution by phylogenetic tree analysis. BMC Bioinformatics 8: 83.
- 17. Vilella AJJ, Severin J, Ureta-Vidal A, Durbin R, Heng L, et al. (2009) Ensemblcompara genetrees: Analysis of complete, duplication aware phylogenetic trees in vertebrates. Genome research 19: 327–335.
- 18. Wapinski I, Pfeffer A, Friedman N, Regev A (2007) Natural history and evolutionary principles of gene duplication in fungi. Nature 449: 54–61.
- 19. Huerta-Cepas J, Bueno A, Dopazo J, Gabaldón T (2008) Phylomedb: a database for genome-wide collections of gene phylogenies. Nucleic Acids Res 36: D491–D496.
- 20. Chen F, Mackey AJ, Stoeckert CJ, Roos DS (2006) Orthomcl-db: querying a comprehensive multispecies collection of ortholog groups. Nucleic Acids Res 34: D363–D368.
- 21. Ostlund G, Schmitt T, Forslund K, Köstler T, Messina DN, et al. (2010) Inparanoid 7: new algorithms and tools for eukaryotic orthology analysis. Nucleic Acids Res 38: D196–D203.
- 22. DeLuca TF, Wu IH, Pu J, Monaghan T, Peshkin L, et al. (2006) Roundup: a multi-genome repository of orthologs and evolutionary distances. Bioinformatics 22: 2044–2046.
- 23. Jothi R, Zotenko E, Tasneem A, Przytycka TM (2006) Coco-cl: hierarchical clustering of homology relations based on evolutionary correlations. Bioinformatics 22: 779–788.
- 24. Dalquen DA, Anisimova M, Gonnet GH, Dessimoz C (2012) ALF–A Simulation Framework for Genome Evolution. Molecular Biology and Evolution 29: 1115–1123.
- 25. Yang Z, Nielsen R, Goldman N, Pedersen AMK (2000) Codon-substitution models for heterogeneous selection pressure at amino acid sites. Genetics 155: 432–449.
- 26. Sayers EW, Barrett T, Benson DA, Bryant SH, Canese K, et al. (2009) Database resources of the national center for biotechnology information. Nucleic acids research 37: D5–15.
- 27. Linard B, Thompson J, Poch O, Lecompte O (2011) Orthoinspector: comprehensive orthology analysis and visual exploration. BMC Bioinformatics 12: 11.
- 28. Trachana K, Larsson TA, Powell S, Chen WH, Doerks T, et al. (2011) Orthology prediction methods: a quality assessment using curated protein families. BioEssays 33: 769–780.
- 29. Dessimoz C, Boeckmann B, Roth A, Gonnet GH (2006) Detecting non-orthology in the cog database and other approaches grouping orthologs using genome-specific best hits. Nucleic Acids Res 34: 3309–3316.
- 30. Gonnet GH, Hallett MT, Korostensky C, Bernardin L (2000) Darwin v. 2.0: An interpreted computer language for the biosciences. Bioinformatics 16: 101–103.
- 31. Kristensen DM, Kannan L, Coleman MK, Wolf YI, Sorokin A, et al. (2010) A low-polynomial algorithm for assembling clusters of orthologous groups from intergenomic symmetric best matches. Bioinformatics 26: 1481–1487.
- 32.
Neyman J (1971) Molecular studies of evolution: a source of novel statistical problems. In: Gupta S, Yackel J, editors, Statistical decision theory and related topics, Academic Press, New York. 1–27.
- 33. Tarjan RE, van Leeuwen J (1984) Worst-case analysis of set union algorithms. Journal of the ACM 31: 245–281.
- 34. Ford L, Fulkerson D (1956) Maximal ow through a network. Canadian Journal of Mathematics 8: 399–404.
- 35. Karger DR, Stein C (1996) A new approach to the minimum cut problem. J ACM 43: 601–640.