Bioinformatics, 31(14), 2015, 2409–2411

doi: 10.1093/bioinformatics/btv161
Advance Access Publication Date: 19 March 2015
Applications Note

Systems biology

MAGNA11: Maximizing Accuracy in Global

Network Alignment via both node and edge

Downloaded from https://academic.oup.com/bioinformatics/article/31/14/2409/256266 by guest on 12 October 2021

conservation
V. Vijayan1, V. Saraph2 and T. Milenković1,*
1
Department of Computer Science and Engineering, ECK Institute for Global Health, Interdisciplinary Center for
Network Science and Application, University of Notre Dame, IN 46556, USA and 2Department of Computer
Science, Brown University, Providence, RI 02912, USA
*To whom correspondence should be addressed.
Associate Editor: Jonathan Wren
Received on November 20, 2014; revised on January 31, 2015; accepted on March 14, 2015

Abstract
Motivation: Network alignment aims to find conserved regions between different networks.
Existing methods aim to maximize total similarity over all aligned nodes (i.e. node conservation).
Then, they evaluate alignment quality by measuring the amount of conserved edges, but only after
the alignment is constructed. Thus, we recently introduced MAGNA (Maximizing Accuracy in
Global Network Alignment) to directly maximize edge conservation while producing alignments
and showed its superiority over the existing methods. Here, we extend the original MAGNA with
several important algorithmic advances into a new MAGNAþþ framework.
Results: MAGNAþþ introduces several novelties: (i) it simultaneously maximizes any one of three
different measures of edge conservation (including our recent superior S3 measure) and any
desired node conservation measure, which further improves alignment quality compared with
maximizing only node conservation or only edge conservation; (ii) it speeds up the original
MAGNA algorithm by parallelizing it to automatically use all available resources, as well as by
reimplementing the edge conservation measures more efficiently; (iii) it provides a friendly graph-
ical user interface for easy use by domain (e.g. biological) scientists; and (iv) at the same time,
MAGNAþþ offers source code for easy extensibility by computational scientists.
Availability and implementation: http://www.nd.edu/cone/MAGNAþþ/
Contact: tmilenko@nd.edu

1 Introduction
Proteins produced by genes interact to carry out cellular processes,
which can be modeled by protein–protein interaction (PPI) net-
works. PPI network alignment can be used to find a node mapping
between networks of different species that identifies similar regions
between the networks (Sharan and Ideker, 2006; Clark and Kalita,
2014). Consequently, it can be used to predict protein function by
transferring knowledge from the network of a well-studied species
to the network of a poorly studied species, or to reconstruct species’
phylogenetic relationships based on similarities of their networks
(Liao et al., 2009; Kuchaiev et al., 2010; Milenković et al., 2010;
Kuchaiev and Pržulj, 2011; Patro and Kingsford, 2012; Faisal et al.,
2014; Saraph and Milenković, 2014; Sun et al., 2014).
Traditionally, existing methods first compute pairwise node simi-
larities between networks and then they find a high-scoring align-
ment that maximizes (greedily or optimally) the total similarity over
all aligned nodes (or node conservation) (Faisal et al., 2014;
Crawford et al., 2014). However, alignment quality is then evaluated
with respect to a measure of edge conservation. Thus, traditional
methods aim to conserve edges by aligning nodes that are similar.

C The Author 2015. Published by Oxford University Press. All rights reserved. For Permissions, please e-mail: journals.permissions@oup.com
V 2409
2410 V.Vijayan et al.

In contrast, our recent MAGNA (Maximizing Accuracy in the accuracy of the alignment in terms of node correctness (NC)
Global Network Alignment) (Saraph and Milenković, 2014) directly (Fig. 1) (Saraph and Milenković, 2014).
maximizes edge conservation while producing alignments. MAGNA
was shown to outperform the existing methods, including IsoRank
2.2 Speedup via parallelization and faster calculation of
(Singh et al., 2007), MI-GRAAL (Kuchaiev and Pržulj, 2011) and
GHOST (Patro and Kingsford, 2012), in terms of both node and
edge conservation
The original MAGNA is a genetic algorithm that combines existing
edge conservation. Importantly, in addition to constructing its own
‘parent’ alignments into superior ‘children’ alignments and then
superior alignments from scratch, MAGNA can combine alignments
evolves this process over multiple generations. Its main computa-
of existing methods to further improve them.
tional bottleneck is the calculation of quality for each alignment in
As simultaneously maximizing both node and edge conservation
each generation, which is needed in order to select the highest-
could further improve alignment quality (Neyshabur et al., 2013;
scoring alignments for the next generation. As quality of each align-
Crawford and Milenković, 2014; Sun et al., 2014), here we extend
ment can be calculated independently, unlike the original single-
MAGNA into a new MAGNAþþ framework, which (i) allows for

Downloaded from https://academic.oup.com/bioinformatics/article/31/14/2409/256266 by guest on 12 October 2021

maximizing a combination of any node conservation (i.e. similarity)
measure and any one of three currently implemented edge conserva-
tion measures; (ii) parallelizes the original MAGNA algorithm via
multi-threading and reimplements MAGNA’s edge conservation
measures more efficiently to decrease running time; (iii) implements
a friendly graphical user interface (GUI) for easy use by domain sci-
entists; and (iv) makes available the source code for easy extensibil-
ity by computational scientists.
2 MAGNA11 details
2.1 Simultaneous node and edge conservation
Given node and edge conservation measures SN and SE, respectively,
MAGNAþþ maximizes aSE þ ð1  aÞSN , where a is a parameter
between 0 and 1 that controls for the contribution of each of the
two measures. Importantly, when we repeat the same analyses as in
the original MAGNA paper, we show that maximizing both node
and edge conservation always improves both node and edge align-
ment quality compared with optimizing only node conservation
(as the existing methods do) or only edge conservation (as the ori-
ginal MAGNA does) (Fig. 1). Here, we have used the total graphlet
degree vector similarity (GDV-similarity) over all aligned nodes
(Milenković and Pržulj, 2008; Milenković et al., 2010) as a measure
of node conservation (but other measures, such as protein sequence
similarity, can also be used), and we have used our recent superior
S3 (Saraph and Milenković, 2014) as a measure of edge conserva-
tion. We have aligned (i) a high-confidence yeast (Y) PPI network
with its five lower-confidence (LC) counterparts, (ii) the same yeast
(Y) network with its five randomly rewired (RW) counterparts, and
(iii) three pairs of real-world PPI networks of different species; for
(i) and (ii), we know the true node mapping and so we can calculate
thread MAGNA, MAGNAþþ divides the calculation among mul-
tiple threads in a way that achieves a dramatic (up to linear) speedup
(Fig. 2a). Parallelizing MAGNAþþ required changes in the imple-
mentation of the original MAGNA in order to account for data
structures that are appropriate for achieving the speedup. Namely,
using adjacency matrices rather than adjacency lists is faster when
dealing with a single thread. However, we found that simply paral-
lelizing the adjacency matrix-based implementation did not realize
high speedup, because of non-trivial cache contention issues with
adjacency matrices when using multiple threads. Thus, to achieve
high enough speedup with multiple threads (Fig. 2a), MAGNAþþ
implementation instead uses adjacency lists.
We achieve additional speedup. Time complexity of calculating
the edge-based alignment quality measures in the original MAGNA,
by iteratively exploring jE1 j edges in the smaller network and a sub-
set of jE2 j edges in the larger network that participate in the aligned
nodes, is OðjE1 jlog jE2 jÞ (Saraph and Milenković, 2014). In
MAGNAþþ, we formulate a new and faster method to calculate
this. Namely, we create a composite graph on jE1 j edges from the
smaller network and the same subset of jE2 j edges from the larger
network and then simply count the number of conserved edges (Fig.
2b). This decreases the above time complexity to OðjE1 j þ jE2 jÞ. We
validate this speedup empirically as well (Fig. 2c).
2.3 User-friendly GUI plus source code
MAGNAþþ has a friendly and intuitive GUI for Windows, Mac OS
X and Unix, plus a command line interface (see its website for a tu-
torial). The only required inputs are the two networks to be aligned
and the output file name. To maximize a node conservation measure
in addition to the default S3 edge conservation measure, the user can
specify the a parameter value and the file with pairwise node simi-
larities (with respect to the desired node conservation measure). The
GUI also allows for advanced options dealing with parallelization of

Y− Y− Y− Y− Y− Y− Y− Y− Y− Y− C. jejuni − Meso. − Yeast −

Y5%LC Y10%LC Y15%LC Y20%LC Y25%LC Y5%RW Y10%RW Y15%RW Y20%RW Y25%RW E. coli Syne. Human
Alignment quality score

2.8
2.4 S3
2 NM
1.6 NC
1.2
0.8
0.4
0
0 0.4 1 0 0.6 1 0 0.6 1 0 0.6 1 0 0.6 1 0 0.4 1 0 0.6 1 0 0.6 1 0 0.8 1 0 0.8 1 0 0.6 1 0 0.4 1 0 0.4 1
α α α α α α α α α α α α α

Fig. 1. Alignment quality in terms of NC, GDV-similarity node conservation measure (NM) and S3 edge conservation measure (S3 ), when optimizing with
MAGNAþþ node conservation only (left; a ¼ 0), edge conservation only (right; a ¼ 1; the original MAGNA), or a combination of node and edge conservation (mid-
dle; a in the (0, 1) range). We show results for the same 13 synthetic and real-world network pairs (shown at the top of the figure) as in the original MAGNA publi-
cation (Saraph and Milenković, 2014). Recall that we can compute NC for the synthetic but not real-world networks
Accuracy in global network alignment
Fig. 2. The speedup of MAGNAþþ. (a) The parallelization speedup is shown
as a function of the number of threads (cores), when MAGNAþþ is run on a
64-core machine, for the three real-world network pairs and the five synthetic
yeast (Y) low-confidence (LC) network pairs from Figure 1. We compute
speedup as the ratio of the time to run the process with one thread (the ori-
ginal MAGNA’s time) to the time to run the process with k threads. (b) Given
two networks and their alignment (i.e. node mapping indicated by broken
arrows) shown at the top, their composite graph is shown at the bottom.
Using this composite graph idea allows for achieving additional speedup via
faster calculation of edge conservation. In this graph, double edges (indicated
with presence of both a blue edge and a red edge between the same aligned
node pairs) are conserved under the given alignment. These edges, along
with their participating nodes, form the common conserved subgraph be-
tween the networks. One of outputs of MAGNAþþ is this common conserved
subgraph, which allows one to visualize the given alignment in an intuitive
manner. (c) Comparison of new MAGNAþþ’s and original MAGNA’s S3 im-
plementations is shown in terms of running time (in seconds), with respect to
the same evaluation test, when both MAGNAþþ and MAGNA are fairly run
as single threads, for the same networks as in panel (a). Clearly, for each pair
of networks, the new S3 implementation in MAGNAþþ is always faster than
the original S3 implementation in MAGNA
the genetic algorithm. The output consists of the alignment (the list
of aligned nodes from the input networks), the statistics regarding
the alignment quality and the common conserved subgraph
(Fig. 2b), which can then be analyzed in any network visualization
tool for an intuitive interpretation of the alignment. For details, see
the MAGNAþþ tutorial. MAGNAþþ also provides the source
code. Thus, computational scientists can easily extend it to, for ex-
ample, modify MAGNA’s alignment crossover function (which is at
the heart of its genetic algorithm) or add additional node or edge
conservation measures.
2411
3 Conclusion
MAGNA is an already proven network aligner. MAGNAþþ is its
novel extension that allows for higher alignment quality, lower com-
putational complexity, intuitive use by domain scientists and easy
functional extensibility by computational scientists.
Funding
This work was supported by the National Science Foundation
[CCF-1319469].
Conflict of Interest: none declared.
Downloaded from https://academic.oup.com/bioinformatics/article/31/14/2409/256266 by guest on 12 October 2021
References
Clark,C. and Kalita,J. (2014) A comparison of algorithms for the pairwise
alignment of biological networks. Bioinformatics, 30, 2351–2359.
Crawford,J. and Milenković,T. (2014). GREAT: GRaphlet Edge-based net-
work AlignmenT. arXiv:1410.5103 [q-bio.MN].
Crawford,J. et al. (2014) Fair evaluation of global network aligners.
arXiv:1407.4824 [q-bio.MN].
Faisal,F. et al. (2014) Global network alignment in the context of aging.
Computational Biology and Bioinformatics, IEEE/ACM Trans. Comput.
Biol. Bioinform, 12, 40–52.
Kuchaiev,O. and Pržulj,N. (2011) Integrative network alignment reveals large
regions of global network similarity in yeast and human. Bioinformatics,
27, 1390–1396.
Kuchaiev,O. et al. (2010) Topological network alignment uncovers biological
function and phylogeny. J. R. Soc. Interface, 7, 1341–1354.
Liao,C. et al. (2009) IsoRankN: spectral methods for global alignment of mul-
tiple protein networks. Bioinformatics, 25, i253–i258.
Milenković,T. and Pržulj,N. (2008) Uncovering biological network function
via graphlet degree signatures. Cancer Inform., 6, 257–273.
Milenković,T. et al. (2010) Optimal network alignment with graphlet degree
vectors. Cancer Inform., 9, 121–137.
Neyshabur,B. et al. (2013) NETAL: a new graph-based method for global
alignment of protein-protein interaction networks. Bioinformatics, 29,
1654–1662.
Patro,R. and Kingsford,C. (2012) Global network alignment using multiscale
spectral signatures. Bioinformatics, 28, 3105–3114.
Saraph,V. and Milenković,T. (2014) MAGNA: maximizing accuracy in global
network alignment. Bioinformatics, 30, 2931–2940.
Sharan,R. and Ideker,T. (2006) Modeling cellular machinery through biolo-
gical network comparison. Nat. Biotechnol., 24, 427–433.
Singh,R. et al. (2007) Pairwise global alignment of protein interaction net-
works by matching neighborhood topology. In: Speed,T. and Huang,H.
(eds), Research in Computational Molecular Biology. Springer, Berlin,
Germany, pp. 16–31.
Sun,Y. et al. (2014) Simultaneous optimization of both node and edge
conservation in network alignment via WAVE. arXiv:1410.3301
[q-bio.MN].