A fast approach to global alignment of proteinprotein interaction networks
 Giorgos Kollias^{1, 3}Email author,
 Madan Sathe^{2}Email author,
 Shahin Mohammadi^{1, 3}Email author and
 Ananth Grama^{1, 3}Email author
DOI: 10.1186/17560500635
© Kollias et al.; licensee BioMed Central Ltd. 2013
Received: 11 September 2012
Accepted: 17 January 2013
Published: 31 January 2013
Abstract
Background
Global network alignment has been proposed as an effective tool for computing functional orthology. Commonly used global alignment techniques such as IsoRank rely on a twostep process: the first step is an iterative diffusionbased approach for assigning similarity scores to all possible node pairs (matchings); the second step applies a maximumweight bipartite matching algorithm to this similarity score matrix to identify orthologous node pairs. While demonstrably successful in identifying orthologies beyond those based on sequences, this twostep process is computationally expensive. Recent work on computation of nodepair similarity matrices has demonstrated that the computational cost of the first step can be significantly reduced. The use of these accelerated methods renders the bipartite matching step as the dominant computational cost. This motivates a critical assessment of the tradeoffs of computational cost and solution quality (matching quality, topological matches, and biological significance) associated with the bipartite matching step. In this paper we utilize the stateoftheart core diffusionbased step in IsoRank for similarity matrix computation, and couple it with two heuristic bipartite matching algorithms – a matrixbased greedy approach, and a tunable, adaptive, auctionbased matching algorithm developed by us. We then compare our implementations against the performance and quality characteristics of the solution produced by the reference IsoRank binary, which also implements an optimal matching algorithm.
Results
Using heuristic matching algorithms in the IsoRank pipeline exhibits dramatic speedup improvements; typically ×30 times faster for the total alignment process in most cases of interest. More surprisingly, these improvements in compute times are typically accompanied by better or comparable topological and biological quality for the network alignments generated. These measures are quantified by the number of conserved edges in the alignment graph, the percentage of enriched components, and the total number of covered Gene Ontology (GO) terms.
Conclusions
We have demonstrated significant reductions in global network alignment computation times by coupling heuristic bipartite matching methods with the similarity scoring step of the IsoRank procedure. Our heuristic matching techniques maintain comparable – if not better – quality in resulting alignments. A consequence of our work is that networkalignment based orthologies can be computed within minutes (as compared to hours) on typical protein interaction networks, enabling a more comprehensive tuning of alignment parameters for refined orthologies.
Background
The description of a cell as a collection of pathways of interacting biochemical components is fundamental to a systems view of biological processes. Data relating to regulatory, metabolic, and signaling interactions, is systematically, and naturally encoded into networks [1]. The diversity of species, cellular processes, abstractions, and volume of interaction data generated from highthroughput techniques strongly motivates development of effective and efficient analysis algorithms. Over the past two decades, significant progress has been made on algorithms for identifying conserved components, discriminating components, modularity, clustering, and alignment, of sequences, sets, and special graph structures (trees, DAGs). However, solving these problems for general large sparse graphs, while providing sound statistical basis for results, remains a topic of significant ongoing investigation. Effective solutions to these problems must leverage properties of specific datasets to deliver desirable performance.
In this paper, we focus on the problem of network alignment. This problem aims to quantify the similarity of two given graphs, resulting in the mapping of nodes from one graph to the other, along with the induced edge mapping. As a specific instance of this problem, in protein interaction (PPI) networks, physically interacting proteins are represented as edgeconnected nodes. Identifying (topological) regions of similarity between networks of different species reveals insights into the functional organization and coherence of subnetworks. Specifically, if connected subgraphs are conserved across species, they likely correspond to shared function across and within subgraphs. This can be used for annotating proteins (by mapping annotations across species), inferring missing interactions, and drawing functional orthologies.
Network alignments can be derived from local or global measures of cost. In local network alignment (LNA) [2–4], local scoring functions reward potentially small subgraph matches. In PPIs, it follows that a node (protein) may potentially participate in many mappings. In contrast, global network alignment (GNA) [5–7] relies on a cost function defined over entire networks. This implies that in PPIs, a single protein from one network is mapped to a single protein in the other.
Proposed approaches to global alignment typically proceed in two steps:

In the first step, a similarity score matrix X is constructed, where element x_{ i j } denotes the similarity of node i in the first graph to node j in the second graph.

In the second step, a node matching algorithm selects pairs of nodes, one from each network, optimizing an aggregate similarity score measure using matrix X.
The similarity of two nodes is determined by the similarity of their interaction profiles, also called topological similarity, and the inherent similarity of the nodes. The latter notion of inherent similarity is introduced in by Singh et al. [6, 7], as elemental (or node) similarity. Elemental similarity scores supplement topological similarity, and rely on node labels or attributes. As an example, for protein pairs, sequence similarity scores, independently computed by BLAST, can be used for elemental similarity. The method of Singh et al. [6, 7], called IsoRank, uses a notion of topological similarity based on an iterative diffusion process for computing matrix X. In this method, the similarity of a pair of nodes is iteratively determined by the similarity of their neighbors. The topological similarity thus computed is accumulated with the elemental similarity at each iterative step and the process is run to convergence to yield matrix similarity matrix X.
In the second step of the method, bestmatching pairs of nodes are identified. The matrix X is viewed as a weighted bipartite graph G=(V_{ A },V_{ B },E), where node i of the first graph represents a vertex in V_{ A }, node j of the second graph illustrates a vertex in V_{ B } and matrix entry x_{ i j } represents a weighted edge between the two nodes. Bipartite graph matching algorithms can be applied to obtain a set of edges M, M⊆E, such that no pair of edges in M are incident on/to the same vertex. Furthermore, the sum over the weighted edges in M is maximized (IsoRank [6, 7], NetAlignBP [8], HGRAAL [9]).
In IsoRank, the computation of matrix X represents the dominant cost. However, algorithmic improvements to this step have resulted in significant reductions in the cost of similarity matrix computation [10], particularly in cases with a small set of dominant, elemental similarity components. This has resulted in the second, bipartite matching step now becoming the performance bottleneck, and consequently, the focus of performance improvements.
In this paper, we critically examine the need for optimal bipartite matching, suitable heuristic algorithms for bipartite matching, the associated improvements in overall runtime, and their implications for quality of solution, both in terms of biological implications and topological matches.
Specifically, for a series of PPI network pairs, we experiment with our customized implementation of the two stage pipeline: we utilize the IsoRank scoring matrix calculation as the first stage, followed by the application of the matrixbased, greedy, and adaptive, auctionbased matching algorithms. These two methods generate two alignments for each input network pair – (mat3_greedy and mat3_auction results). We then run, for the same input networks and parameters, the reference native binary implementation of IsoRank made available by Singh et al.[6], who report that their software applies the greedy and Hungarian algorithms to the scoring matrix X, and produces iso_greedy and iso_hungarian alignments.
Our key result is an improvement in overall runtime of over one order of magnitude – typically an acceleration of ×30for almost all tested PPI network pairs – with comparable, and in some cases superior results in terms of topological and biological quality (mat3_* versus iso_* computations). This represents a reduction in runtime for typical cases of PPI networks from hours to minutes – thus enabling a more comprehensive exploration of the parameter space for extracting desirable orthologies.
Results and discussion
PPI network data
Species  Dataset name   V   E 

fruitfly(D. melanogaster)  dmela  7518  25830 
bacterium(E. coli)  ecoli  1821  6849 
human(H. sapiens)  hsapi  9633  36386 
yeast(S. cerevisiae)  scere  5499  31898 
Computational performance
Two important observations can be made:

Our mat3_* matches are generated approximately 30 times faster than iso_*, for almost all pairs of networks. For the largest datasets this roughly translates to less than 5 minutes in our case (compared to iso_*’s 2 hours of processing time).

The adaptive auction algorithm is a key element in this performance improvement, as opposed to the Hungarian algorithm — assuming IsoRank (constructing X) and greedy matching implementations are of comparable performance. Please note that we have no way of benchmarking these separately, since these times are not reported by the reference implementation of Singh et al. One may argue that our implementations of similarity matrix computations are much faster than corresponding implementations of Singh et al. While this is unlikely, even if this were to be the case, our significant performance improvements over the most widely distributed implementation of IsoRank represent the core of our contributions.
We also note that, especially for the largest pair of networks, our adaptive auction algorithm is the most efficient method for extracting matching pairs from similarity score matrices. Specifically, the only timing information that is available after running the native binary from [11] is the overall running time, which includes three separate components – (i) their similarity matrix construction phase, (ii) their Hungarian algorithm run, and (iii) their greedy approach. These three times cannot be separated. However we can measure, separately, the times for (i) our similarity matrix construction phase, (ii) our auction matching run, and (iii) our greedy approach run. Only the implementation of the second part, i.e, their Hungarian versus our auction algorithm, are fundamentally different. The corresponding similarity computations in the first phase and the greedy matching algorithm rely on similar algorithms. This suggests that the performance gain that we observe in the overall running time from the three parts in both cases (roughly ×30 speedup) can be primarily attributed to the performance difference between their Hungarian and our auction algorithm implementations.
In terms of the number of operations, the complexity of the Hungarian algorithm is $\mathcal{O}\left(\rightV\left\right(\leftE\right+\leftV\rightlog\leftV\right\left)\right)$. This is in comparison to the worst case complexity of an ϵscaling auction algorithm, which is given by $\mathcal{O}\left(\rightV\left\rightElog(\leftV\rightC\left)\right)$ for integer weights (with C= maxi jx_{ i j } and x_{ i j } the similarity scores). Here V is the maximum number of vertices between the two graphs to be aligned and E is the number of entries in X. Experimental results suggest that the worst case runtime for the auction algorithm is rare. Other causes for the speed improvements include suboptimal data structures in similarity matrix computations in the binary code, or an unoptimized Hungarian implementation. These are hard to decipher from the binary – in either case, our software yields significant overall acceleration of the stateoftheart approach from Singh et al. for the global alignment problem.
The alignment graph and its assessment
Topological perspective
In Figure 2, the number of conserved edges in the alignment graph is reported. Our proposed mat3_auction approach outperforms the other methods in at least 2 out of the 6 cases – more conserved edges imply better alignments, as described in the Methods Section – and that the mat3_* matchings are superior to the iso_* ones (in 5 out of 6 cases, and on average). However there is no clear “winner”, i.e. a single method that is the best for all test cases.
Biological perspective
Biological validation of alignment graphs
(a) D.melanogaster vs S.cerevisiae  

Fly  Yeast  
Method  TPR  TNR  TPR  TNR 
iso_greedy  535  17.00%  324  43.00% 
iso_hungarian  457  15.20%  1066  44.00% 
mat3_auction  490  15.40%  1132  38.30% 
mat3_greedy  448  17.00%  1102  44.00% 
(b) D.melanogaster vs H.sapiens  
Fly  Human  
Method  TPR  TNR  TPR  TNR 
iso_greedy  1067  30.10%  191  9.00% 
iso_hungarian  974  31.50%  671  14.30% 
mat3_auction  1020  27.00%  519  12.50% 
mat3_greedy  1029  30.80%  670  17.00% 
(c) D.melanogaster vs E.coli  
Fly  Bacterium  
Method  TPR  TNR  TPR  TNR 
iso_greedy  41  11.63%  66  16.60% 
iso_hungarian  60  10.38%  175  29.35% 
mat3_auction  90  9.80%  236  32.00% 
mat3_greedy  56  10.73%  235  32.58% 
(d) H.sapiens vs S.cerevisiae  
Human  Yeast  
Method  TPR  TNR  TPR  TNR 
iso_greedy  971  31.74%  406  17.97% 
iso_hungarian  1010  31.27%  1063  41.76% 
mat3_auction  1064  28.40%  1083  40.16% 
mat3_greedy  989  33.33%  1095  46.14% 
(e) S.cerevisiae vs E.coli  
Yeast  Bacterium  
Method  TPR  TNR  TPR  TNR 
iso_greedy  419  46.10%  63  15.00% 
iso_hungarian  359  46.60%  231  38.24% 
mat3_auction  400  40.72%  274  35.75% 
mat3_greedy  378  44.22%  289  35.37% 
(f) E.coli vs H.sapiens  
Bacterium  Human  
Method  TPR  TNR  TPR  TNR 
iso_greedy  52  16.77%  472  29.56% 
iso_hungarian  158  35.63%  386  28.40% 
mat3_auction  252  32.96%  557  34.80% 
mat3_greedy  219  36.00%  444  29.73% 
Sensitivity
Specificity
In terms of average specificity, we observe more diversity among different methods. Surprisingly, mat3_greedy is the topranked method in 4 out of 6 experiments, except in aligning yeastversusbacterium and bacteriumversushuman, for which iso_hungarian and mat3_auction perform better, respectively. These results suggest that the wellknown Hungarian algorithm for maximum weighted bipartite matching does not necessarily enhance the biological quality of the results. One possible explanation for this phenomena is the overfitting problem. The objective function for the matching phase is defined over the set of pairwise similarity scores for the nodes in different graphs, which itself is computed using both sequence and topological similarities. We note that both of the initial scores – sequence similarity scores computed using BLAST, and aggregated scores computed using IsoRank, are inherently noisy and overfitting a model (matching) on them can potentially decrease the performance of the results.
Examples of highly enriched components
These functionally similar modules exhibit closely related structure. In most cases, the smaller subgraph and all its edges can be perfectly embedded in the other species. The set of missed interactions can be explained by either evolutionary divergence or the quality of the input dataset.
From an evolutionary standpoint, one can argue that the set of orthologous genes in the pair of species could have diverged and either gained or lost specific functions. These functional adaptations are reflected in the pattern of proteinprotein interactions by deleting or inserting partnerships. On the other hand, an easier argument can be made based on the quality and the coverage of proteome, noting that different PPI interactions have been predicted in different labs for each species. Some of these species are wellstudied and there are more highquality interactions available for them. As is evident from Figures 3(a) and 3(d), most of the missed edges reside in yeast, which can simply be explained by heterogeneity in the quality of initial PPI networks.
Conclusions
Our results show that the IsoRankbased method for computing similarity scores between nodes of two PPI networks, coupled with a fast, adaptive, auctionbased implementation and a matrixbased greedy algorithm for extracting matching proteins, yields an efficient and effective algorithm for global alignment. The method yields over one order of magnitude improvement in computation time – typically ×30 times faster for the total alignment process in most cases of interest – with comparable or superior topological and biological quality of the results. Using out method, alignments can be computed within minutes (less than 5 minutes for typical alignments), as compared to hours, using previously used methods (of the order of 130 minutes for the largest input configurations). This enables users to tune alignment parameters much better and extract superior alignments.
Methods
Ranking node pairs in IsoRank
We use the similarity matrix construction step in IsoRank in all alignment methods. This method is based on an analogy between the network alignment problem and that of identifying “reputed” nodes in a single network – also sometimes called the page or noderanking problem. Perhaps, the most commonly used measure for the rank of a node in a single given network can be recursively defined as follows: “a node is important if it is linked to other important nodes” [12]. Extending this definition to our node similarity problem, we arrive at the following definition: “two nodes are (topologically) similar if they are linked to other (topologically) similar node pairs” [6, 7, 13]. IsoRank effectively implements this notion of “recursive node similarity”. Note that this perspective is not new – it has been exploited in application areas like automated image captioning [14] and synonym extraction [15].
We initiate a more formal discussion by introducing necessary notation. Given a graph G_{ A }=(V_{ A },E_{ A }), V_{ A } and E_{ A } denote the vertices and edges of G_{ A } respectively, and n_{ A }=V_{ A }. Its adjacency matrix A has elements a_{ i j }=1 iff edge (i,j)∈E, and a_{ i j }=0 otherwise. Clearly, A is symmetric for an undirected graph. Matrix $\xc3$ is the normalized version of the matrix A^{ T }; formally, ${\left(\xc3\right)}_{\mathit{\text{ij}}}={a}_{\mathit{\text{ji}}}/\sum _{i=1}^{{n}_{A}}{a}_{\mathit{\text{ji}}}$ for nonzero rows of A and zero otherwise. We denote by 1, the column vector of size n_{ A } consisting of 1’s. Also, the v e c(·) operator for stacking matrix columns into a vector (as well as its associated “inverse” u n v e c(·) operator for reassembling the matrix) are used. Using these operators, v e c(A X B)=(B^{ T }⊗A)x, holds, where ⊗ denotes the Kronecker product of two matrices.
Matching algorithms
As a first step in identifying similar subgraphs in the two networks, IsoRank computes a V_{ A }by V_{ B } similarity score matrix X. We assume without loss of generality that V_{ A }≤V_{ B }. The matrix can be viewed as encoding a weighted bipartite graph G=(V_{ A },V_{ B },E,w), where V_{ A }∩V_{ B }=∅, E⊆V_{ A }×V_{ B }, and the weight function $w:E\to {\mathbb{R}}^{\ge 0}$. Each row represents a vertex in V_{ A } and each column a vertex in V_{ B }. A nonzero entry in the matrix is interpreted as an edge between the row and column vertices. The numerical value x_{ i j } of the matrix indicates the weight of the edge. A matching in the bipartite graph is defined as a subset M⊆E such that no pair of edges of M are incident on the same vertex. In a maximal matching, no edge can be added to M without violating the matching property. A maximum (cardinality) matching is a matching that contains the largest possible number of edges. Specifically, we are interested in finding a maximum matching with maximum weight. The weight of the matching is defined as $\sum _{(i,j)\in M}{x}_{\mathit{\text{ij}}}$; this typically translates to a large number of matching pairs with a high cumulative similarity score.
Various algorithms have been proposed for computing a weighted matching M for a given similarity matrix X. Approximation algorithms [16], which compute a maximal matching with a maximum weight, and maximum weighted matching algorithms [17, 18], which return a maximum matching with a maximum weight, are all candidates for finding a suitable assignment. In our approach, we use the 1/2approximation algorithm and propose a weighted matching implementation that is based on the principle of auctions. The main advantage of using an auctionbased scheme is the socalled ϵscaling mechanism, using which the quality and convergence of the algorithm can be controlled.
1/2approximation algorithm
This simple approximation algorithm can be described as follows: First, the weights of the edges are sorted in decreasing order. Then, the heaviest edge e is selected and deleted, together with the edges incident to its endpoints. This is repeated until the graph is empty. The worst case complexity of this sorting procedure is $\mathcal{O}\left(\rightElogE\left\right)$. We implement the lineartime 1/2approximation algorithm of Preis [16], of known time complexity $\mathcal{O}\left(\rightE\left\right)$ (Algorithm 1), translating this graphbased description into matrix operations (matrixbased greedy algorithm): After selecting the element with maximum value x_{ i j } from the similarity matrix X≠0, we report the matching of nodes i and j. We then zero the i^{ t h } row and the j^{ t h } column of X, and repeat the aforementioned step until X contains only zeros across one of its dimensions.
Algorithm 1 1/2approximation algorithm for weighted matching
Auction algorithms
Auction algorithms [18] find the maximum weighted matching via an auction: in this analogy, i∈V_{ A } is a person, j∈V_{ B } is an object and x_{ i j } is the benefit the buyer i obtains by acquiring object j. Each object j has an associated price p_{ j }, with the initial value zero.
Algorithm 2 Auction Algorithm
In an auction iteration, the bidding and assignment phase, and update of the price and of the value of ϵ are performed. In the bidding phase, an unassigned buyer i bids for the best object j_{ i }, i.e., the object j_{ i }, that has the maximal profit for buyer i. The bid is computed by subtracting the secondbest profit v_{ i } from the most valuable profit u_{ i }, i.e., u_{ i }−v_{ i }. The most valuable profit u_{ i } for buyer i is defined as $\{{x}_{i{j}_{i}}{p}_{{j}_{i}}\}$, while the secondbest profit v_{ i } is computed by $\underset{j\ne {j}_{i}}{max}\{{x}_{\mathit{\text{ij}}}{p}_{j}\}$. After the unassigned buyer has submitted the bid, the designated object is awarded to the bidder, yielding its potential previous owner unassigned. The price is calculated by updating the old price by the corresponding bid and by a small increment ϵ. It follows that the auctionbased algorithm (see Algorithm 2 for a simplified description also assuming integer x_{ i j }) consists of four phases: the initialization phase (lines 1–7), the bidding phase (lines 9–12), the assignment phase (lines 13–15), and the termination phase (line 8).
The initial value of increment ϵ has significant impact on the computational cost of the auction algorithm. Ideally, the value of ϵ should be close to the optimal value of the price, since the number of iterations to find a matching will be small. In general, the computational worst case complexity of an ϵscaling auction algorithm is $\mathcal{O}\left(\rightV\left\rightElog(\leftV\rightC\left)\right)$ (where C is defined in line 5 of Algorithm 2). Due to the pseudopolynomial complexity, we embed an aggressive ϵscaling strategy into the auctionbased implementation of Algorithm 2, resulting in our adaptive auction algorithm, which effectively splits the ϵscaling phase into multiple ϵscaling phases; its steps are detailed in Algorithm 3.
Algorithm 3 Adaptive Auction Algorithm
Here, the ϵvalue is initialized with a small value and adaptively increased relative to the overall progress of the matching. The basic idea behind the proposed heuristic is that in the inner iteration, at least δ buyers get assigned to an object, while ϵ pushes the price of the object to a large value. In general, the heuristic provides a maximal matching with a maximum weight. Since the matrix X is dense, a simple greedy approach is applied subsequently, in order to achieve a maximum matching. The postprocessing method assigns an unmatched buyer to the first unmatched object from the object list, resulting in a maximum matching with maximum weight.
For a more detailed presentation of the auction algorithm we use, we refer readers to Sathe et al. [19], where a scalable distributed version of the algorithm is described. This formulation computes weighted matchings on sparse and dense bipartite graphs running on hundreds of compute nodes, while efficiently using multiple cores on each compute node.
Numerical experiments
In all cases, the construction of the similarity matrix X is IsoRankbased. We implemented the iterative scheme of Equation (2) in Matlab (mat3_* cases — name mat3 is used to indicate a triplematrix product kernel) and tested this part against equivalent codes in the netalign package [20], yielding exact agreement (within machine accuracy). The resulting matrix X of similarity scores is then input to (i) our matrixbased implementation (in Matlab) of the 1/2approximation algorithm (hereafter referred to as greedy) to produce the mat3_greedy alignment and (ii) to the adaptive auction algorithm (in C) to generate the respective mat3_auction alignment.
We then execute, for the same runtime parameters, the reference IsoRank binary for producing iso_greedy and iso_hungarian alignments and compare them against our approach (mat3_* alignments). Unfortunately the reference native code does not provide an option for generating either the similarity matrix X, or the timings for its computation. It internally uses the result of this first phase to extract the best matching node pairs, the second phase. The total timing results — for both phases — are reported together with the extracted matchings pairs in this case. Specifically, results from two matching algorithms are reported [6], namely their implementation of the greedy and Hungarian algorithms – thus justifying the selection of iso_greedy and iso_hungarian names for the complete alignment pipelines. It is understood that marked deviations could arise due to internal optimizationtargeted realizations of the algorithmic ideas, however these are not readily accessible or reported. Figure 4 summarizes the algorithmic blocks driving the generations of four sets of matchings for each pair of PPI networks.
The alignment graph and its assessment
Topological perspective
When analyzing the alignment graph of two networks, two measures for the topological evaluation of the computed match can be used:

The number of edges in the alignment graph (conserved edges). Note that the more the conserved edges – incident on our matching nodes in the two networks – the larger the percentage of “link structure” (edges at minimal or larger connected subgraphs) that could be put under direct mapping as well, i.e. aligned.

The size of the connected components in the alignment graph (common connected subgraphs). These are “clusters” of pairs of matching nodes in the original graphs also conserving their link patterns.
The existence of many conserved edges clearly increases the probability of them being part of extensive connected subgraphs. However, it could also be the case that they are parts of larger numbers of connected subgraphs, all of moderate sizes. Comparative, connected subgraph statistics is the main focus of quality assessment from the biological point of view.
Biological perspective
Topological assessment of the alignment graph can uncover important characteristics of each alignment method. However, additional considerations must be taken into account to avoid growing components at the expense of overgeneralizing them. This can negatively affect the specificity of predictions. For example, by analyzing the size of connected components, one can penalize against fragmenting functional modules. An alignment with many singletons (isolated pairs of nodes in the alignment graph) is less desirable than a larger connected component that can group a number of related nodes (and their corresponding proteins) together. On the other hand, by mixing functionally independent groups that share a small subset of genes, one can create larger components that are not functionally coherent. To remedy this problem, we adopt an approach similar to the one proposed by Kalaev et al. [21] to assess the functional coherence of each connected component in the alignment graph. In this approach, each connected component is treated as a computational prediction of a functionally related group of proteins and is crossvalidated against the existing GO annotations as the actual set of functionally related genes. Given the gene ontology (GO) [22] annotations with respect to biological process (BP) for the member genes of each species, we compute the set of enriched GO terms within each connected component in the alignment graph using the GO::TermFinder tool [23]. We process each graph separately and, similar to Kalaev et al. [21], apply a threshold of 0.05 to extract the set of enriched GO terms. Finally, we define two complementary criteria to validate each alignment method – the fraction of enriched components (components with at least one enriched GO term) and the total number of covered GO terms in all components. The former captures specificity (true negative rate or TNR), while the latter captures the concept of sensitivity (true positive rate or TPR).
Source code and datasets
All online material for this project is available at http://compbio.soihub.org/projects/fastalign/. This includes Matlab scripts, C code for the (more general, distributed version of the) auction algorithm, input datasets, and also data generated during intermediate phases of an actual run.
Abbreviations
 PPI:

Protein Protein Interaction
 DAG:

Directed Acyclic Graph
 GO:

Gene Ontology
 LNA:

Local Network Alignment.
Declarations
Acknowledgements
This work was supported by the Center for Science of Information (CSoI), an NSF Science and Technology Center, under grant agreement CCF0939370, and by NSF grants DBI 0835677 and 0641037.
Authors’ Affiliations
References
 Palsson B: Systems Biology: Properties of Reconstructed Networks. 2006, Cambridge, United Kingdom: Cambridge University Press,http://www.cambridge.org/gb/knowledge/isbn/item1173392/?site_locale=en_GB,View ArticleGoogle Scholar
 Kelley B, Yuan B, Lewitter F, Sharan R, Stockwell B, Ideker T: PathBLAST: a tool for alignment of protein interaction networks. Nucleic Acids Res. 2004, 32 (suppl 2): W83.PubMedPubMed CentralView ArticleGoogle Scholar
 Sharan R, Suthram, et al.: Conserved patterns of protein interaction in multiple species. Proc Natl Acad Sci. 1974, 102 (6):
 Koyutürk M, Kim Y, Topkara U, Subramaniam S, Szpankowski W, Grama A: Pairwise alignment of protein interaction networks. J Comput Biol. 2006, 13 (2): 182199. 10.1089/cmb.2006.13.182.PubMedView ArticleGoogle Scholar
 Pržulj N, Milenković T: Uncovering biological network function via graphlet degree signatures. Cancer Inform. 2008, 6: 257273.PubMedPubMed CentralGoogle Scholar
 Singh R, Xu J, Berger B: Pairwise global alignment of protein interaction networks by matching neighborhood topology. RECOMB, Volume 4453 of Lecture Notes in Computer Science. Edited by: Speed TP, Huang H. 2007, : Springer, 1631. [http://dx.doi.org/10.1007/9783540716815_2]Google Scholar
 Singh R, Xu J, Berger B: Global alignment of multiple protein interaction networks with application to functional orthology detection. Proc Natl Acad Sci. 1276, 105 (35): 312768.Google Scholar
 Bayati M, Gerritsen M, Gleich D, Saberi A, Wang Y: Algorithms for large, sparse network alignment problems. Data Mining, 2009 ICDM’09. Ninth IEEE International, Conference on. 2009, : , 705710. 10.1109/ICDM.2009.135.View ArticleGoogle Scholar
 Ng W, Hayes W, Pržulj N, Milenković T: Optimal network alignment with graphlet degree vectors. Cancer Inform. 2010, 9: 121PubMedPubMed CentralGoogle Scholar
 Kollias G, Mohammadi S, Grama A: Network Similarity Decomposition (NSD): A fast and scalable approach to network alignment. Knowl Data Eng, IEEE Trans. 2012, 24 (12): 22322243.View ArticleGoogle Scholar
 Rohit Singh and BonnieBerger: IsoRank and IsoRankN. http://groups.csail.mit.edu/cb/mna/[Datasets and IsoRankbinary retrieved in June 2010]
 Page L, Brin S, Motwani R, Winograd T: The PageRank citation ranking: bringing order to the web. Tech. rep., Stanford University 1998
 Jeh G, Widom J: SimRank: a measure of structuralcontext similarity. Proceedings of the eighth ACM SIGKDD international conference on Knowledge discovery and data mining. 2002, KDD ’02. New York: ACM, 538543. [http://doi.acm.org/10.1145/775047.775126]View ArticleGoogle Scholar
 Pan JY, Yang HJ, Faloutsos C, Duygulu P: Automatic multimedia crossmodal correlation discovery. Proceedings of the tenth ACM SIGKDD international conference on Knowledge discovery and data mining. 2004, KDD ’04 New York: ACM, 653658. [http://doi.acm.org/10.1145/1014052.1014135]Google Scholar
 Blondel V, Gajardo A, Heymans M, Senellart P, Van Dooren P: A measure of similarity between graph vertices: applications to synonym extraction and web searching. SIAM Rev. 2004, 46 (4): 647666. 10.1137/S0036144502415960.View ArticleGoogle Scholar
 Preis R: Linear time 1/2 approximation algorithm for maximum weighted matching in general graphs. Proceedings of the 16th annual conference on Theoretical aspects of computer science. 1999, STACS’99. Berlin, Heidelberg: SpringerVerlag, 259269. [http://dl.acm.org/citation.cfm?id=1764891.1764924]Google Scholar
 Kuhn HW: The Hungarian Method for the assignment problem. Naval Res Logistics Q. 1955, 2: 8397. 10.1002/nav.3800020109.View ArticleGoogle Scholar
 Bertsekas DP: The auction algorithm: A distributed relaxation method for the assignment problem. Ann Oper Res. 1988, 14: 105123. 10.1007/BF02186476.View ArticleGoogle Scholar
 Sathe M, Schenk O, Burkhart H: An auctionbased weighted matching implementation on massively parallel architectures. Parallel Comput. 2012, 38 (12): 595614. 10.1016/j.parco.2012.09.001. [http://dblp.unitrier.de/db/journals/pc/pc38.html#SatheSB12]View ArticleGoogle Scholar
 David F Gleich: netalign: Network Alignment codes.http://www.cs.purdue.edu/homes/dgleich/codes/netalign/,
 Kalaev M, Bafna V, Sharan R: Fast and accurate alignment of multiple protein networks. RECOMB. 2008, : , 246256.Google Scholar
 Ashburner M, Ball, et al.: Gene ontology: tool for the unification of biology. The Gene Ontology Consortium. Nature Genet. 2000, 25: 2529. 10.1038/75556.PubMedPubMed CentralView ArticleGoogle Scholar
 Boyle EI, Weng, et al.: GO::TermFinderopen source software for accessing Gene Ontology information and finding significantly enriched Gene Ontology terms associated with a list of genes. Bioinformatics. 2004, 20 (18): 3710+10.1093/bioinformatics/bth456.PubMedPubMed CentralView ArticleGoogle 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.