Construction of reliable protein-protein interaction networks with a new interaction generality measure.

Abstract

MOTIVATION: Recent screening techniques have made large amounts of protein-protein interaction data available, from which biologically important information such as the function of uncharacterized proteins, the existence of novel protein complexes, and novel signal-transduction pathways can be discovered. However, experimental data on protein interactions contain many false positives, making these discoveries difficult. Therefore computational methods of assessing the reliability of each candidate protein-protein interaction are urgently needed. RESULTS: We developed a new 'interaction generality' measure (IG2) to assess the reliability of protein-protein interactions using only the topological properties of their interaction-network structure. Using yeast protein-protein interaction data, we showed that reliable protein-protein interactions had significantly lower IG2 values than less-reliable interactions, suggesting that IG2 values can be used to evaluate and filter interaction data to enable the construction of reliable protein-protein interaction networks.

Full text

Page 1

BIOINFORMATICS
Vol. 19 no. 6 2003, pages 756–763
DOI: 10.1093/bioinformatics/btg070
Construction of reliable protein–protein
interaction networks with a new interaction
generality measure
Rintaro Saito†, Harukazu Suzuki∗and Yoshihide Hayashizaki
Laboratory for Genome Exploration Research Group, RIKEN Genomic Sciences
Center (GSC), 1-7-22 Suehiro-cho, Tsurumi-ku, Yokohama 230-0045, and Genome
Science Laboratory, RIKEN, 2-1 Hirosawa, Wako, Saitama 351-0198, Japan
Received on September 9, 2002; revised on December 3, 2002; accepted on December 6, 2002
ABSTRACT
Motivation: Recent screening techniques have made
large amounts of protein–protein interaction data avail-
able, from which biologically important information such
as the function of uncharacterized proteins, the existence
of novel protein complexes, and novel signal-transduction
pathways can be discovered. However, experimental data
on protein interactions contain many false positives, mak-
ing these discoveries difficult. Therefore computational
methods of assessing the reliability of each candidate
protein–protein interaction are urgently needed.
Results: We developed a new ‘interaction generality’
measure (IG2) to assess the reliability of protein–protein
interactions using only the topological properties of their
interaction-network structure. Using yeast protein–protein
interaction data, we showed that reliable protein–protein
interactions had significantly lower IG2 values than less-
reliable interactions, suggesting that IG2 values can be
used to evaluate and filter interaction data to enable
the construction of reliable protein–protein interaction
networks.
Availability: The protein–protein interaction data used
in this study along with the associated IG2 values are
available at http://genome.gsc.riken.go.jp.
Contact: rgscerg@gsc.riken.go.jp
INTRODUCTION
As whole-genome and complete cDNA sequences became
available for numerous organisms (Adams et al., 2000;
Goffeau et al., 1996; Kawai et al., 2001; Lander et al.,
2001; The C. elegans Sequencing Consortium, 1998;
Venter et al., 2001), the focus of many research efforts
is shifting rapidly from genomics to proteomics. One of
the most important approaches in proteomics is the large-
∗To whom correspondence should be addressed.
†Present address: Institute for Advanced Biosciences, Keio University, 14-1
Baba-cho, Tsuruoka, Yamagata 997-0035, Japan.
scale analysis of protein–protein interactions, because
most proteins function within complexes (Oliver, 2000;
Pawson and Nash, 2000). High-throughput genome-wide
screening for protein–protein interactions has been car-
ried out in yeast, Caenorhabditis elegans, and higher
organisms such as the mouse (Ito et al., 2001; Suzuki
et al., 2001; Uetz et al., 2000; Walhout et al., 2000).
Several successful computational analyses of interaction
data have also been completed (Fellenberg et al., 2000;
Schwikowski et al., 2000).
However, the publicly available protein–protein inter-
action data, especially those obtained from two-hybrid
systems, include many false-positive interactions (Legrain
et al., 2001). Von Mering et al. (2002) estimate that
approximately half the interactions obtained from high-
throughput data may be false positives. These false
positives may unnecessarily link unrelated proteins,
resulting in huge apparent interaction clusters (Ito et al.,
2001), which complicate elucidation of the biological
importance of these interactions. Therefore, a method to
assess the reliability of each candidate protein–protein
interaction is necessary. Earlier, we developed a simple
computational method, which yielded an ‘interaction
generality’ measure (IG1) that could be used to assess
the reliability of experimentally identified interactions
from just a list of interaction data (Saito et al., 2002). The
development of IG1 was based on the idea that interacting
proteins that appear to have many other interacting
partners which have no further interactions are likely to
be false positives. However, our IG1 method was a simple
method for evaluating the reliability of interactions that
did not consider the topological properties of the protein
interaction network beyond the target pair of proteins.
Here we define a new interaction generality (IG2) measure
that overcomes this problem and show that it can assess
the reliability of putative protein–protein interactions with
higher accuracy.
756
Bioinformatics 19(6) c ? Oxford University Press 2003; all rights reserved.
by guest on July 16, 2011
bioinformatics.oxfordjournals.org
Downloaded from

Page 2

Protein–protein interaction generality measure
AB
C
AB
C
AB
C
AB
C
AB
C
a1
a2l
f
d
Fig. 1. Classification of a protein C that interacts with a target
interacting protein pair A–B, according to the topological properties
of its interaction network.
MATERIALS AND METHODS
Preparation of protein–protein interaction data
The publicly available protein interaction data of
Ito et al. (2001); Uetz et al. (2000), and MIPS
(Mewes
etal., 2000)were
//genome.c.kanazawa-u.ac.jp/Y2H/ (754 heterodimers),
http://www.genome.ad.jp/brite/ (905 heterodimers), and
http://www.mips.biochem.mpg.de/proj/yeast/ (2474 het-
erodimers), respectively. We assessed only heterodimers
and considered the interactions of protein A (bait)–protein
B (prey) and of protein B (bait)–protein A (prey) to rep-
resent a single interaction (these represent bidirectional
interactions in a two-hybrid experiment). Combining
these three data sets and removing redundancy from them
yielded 3066 heterodimers, including 673 reproducible
interactions. Interactions that were confirmed by coim-
munoprecipitation assays and bidirectional interactions
that were obtained from two-hybrid assays are considered
to be reproducible.
obtainedfromhttp:
The new interaction generality measure
The new interaction generality measure incorporates the
topological properties of interactions around the target
interacting pairs. The IG2 for the target interacting pair
A–B is defined by the following procedure. Protein C,
which interacts directly with the target interacting pair
A–B, can be classified into one of five groups (a1, a2,
l, f and d) according to the topological properties of its
interaction network (Fig. 1). When C interacts with both
A and B, it is classified as a1 (alternative pathway from
protein A to B through 1 protein). When C interacts with
A but not B, and C also interacts with another protein
that interacts with B, it is classified as a2 (alternative
pathway from protein A to B through 2 protein). When
C is not classified as a2, interacts with A but not B, and
interacts with at least one protein that interacts with A, it
is classified as l (looping interaction). If C does not meet
these three conditions and interacts with another protein,
it is classified as f (further interaction). If C does not
interact with any proteins except for either A or B, it is
classified as d (dead-end interaction).
Then, the numbers of proteins in the database that
belong to each class are counted as n = (Na1, Na2, Nl,
N f , Nd). We counted n’s (n1, n2,...,np, where p is the
number of interactions in the given interaction network)
for all p interactions. From the set of n’s, we constructed
this matrix:


np


Na1p
Na2p
We defined the IG2 value for each interaction by applying
principal-component analysis (Weller and Romney, 1990)
to N in the following way. First, the averages of each
column were subtracted from N, producing


Na1
N =




n1
n2
...
np−1






=




Na11
Na12
...
Na1p−1
Na21
Na22
...
Na2p−1
Nl1
Nl2
...
Nlp−1
Nlp
N f1
N f2
...
N fp−1
N fp
Nd1
Nd2
...
Ndp−1
Ndp






Nc= N − N = N −



Na1
Na1
...
Na2
Na2
...
Na2
Nl
Nl
...
Nl
N f
N f
...
N f
Nd
Nd
...
Nd





where N denotes the average of N and NT
correlations between variables (Na1, Na2, Nl, N f , Nd).
Note that (1/p)NT
cNc represents the covariance matrix
of these variables. Then we determined the matrix P
that satisfied the following equations by singular-value
decomposition (Weller and Romney, 1990):
NT
P = (p1,p2,p3,p4,p5),


000
where piand λiis each eigenvector and its corresponding
eigenvalue of NT
cNcand they satisfy the following equa-
tions:
NT
cNcrepresents
cNc= PDP−1,
D =



λ1
0
0
0
0
λ2
0
0
0
0
λ3
0
0
0
0
λ4
0
0
0
0
0
λ5





cNcpi= λipi(i = 1,2,3,4,5)
λ1> λ2> λ3> λ4> λ5
757
by guest on July 16, 2011
bioinformatics.oxfordjournals.org
Downloaded from

Page 3

R.Saito et al.
Fig. 2. Interaction networks involving the interacting pairs of yeast proteins MED8–MED4 and YMD8–EST1. Nodes and lines denote
proteins and their interactions, respectively. Only proteins within two interaction steps of the target interacting pairs are shown. The
MED8–MED4 interaction was experimentally verified by coimmunoprecipitation (Myers et al., 1998), and the proteins were confirmed
as components of a protein complex (Lorch et al., 2000; Malik and Roeder, 2000), whereas the YMD8–EST1 interaction was indicated only
in a high-throughput two-hybrid assay (Uetz et al., 2000).
Singular-value
tions between the variables (Na1, Na2, Nl, N f ,
Nd), and it summarizes the matrix Nc as vectors pi
(i = 1,2,3,4,5), which are orthogonal to each other.
The λi(i = 1,2,3,4,5) values indicate how well the
matrix is summarized by each corresponding pi. As λ1
has the greatest value among five λ’s, p1is the best vector
that summarizes the matrix Nc. IG2 vectors for each
interaction are defined as follows:
decompositionconsiderscorrela-
IG2 vectors = NcP = (N−N)P = N P−K
IG2 values for each interaction are defined as
(K = N P)
IG2 values = Ncp1= (N − N)p1= Np1− k1
= Na1p11+ Na2p21+ Nlp31+ N f p41
+Ndp51−k1
(k1= Np1,p1= (p11,p21,p31,p41,p51)T)
Thus, the topology of the protein–protein interaction
structure is summarized by a single number IG2. In other
words, characteristics of the topology are mapped to an
IG2 value ∈ R1.
The idea behind the mechanics of IG2 calculation
is that interactions involving proteins that have many
interacting partners are likely to be false positives, but
highly interconnected sets of interactions or interactions
forming a closed loop (such as the set of interaction pairs
A–B, B–C, C–D, D–A) are likely to be true positives
(Walhout et al., 2000; Saito et al., 2002). To distinguish
the false and true interactions, proteins that interact with
the target interaction pair were classified into five classes
(a1, a2, l, f , and d) as mentioned above. Classes a1, a2,
and l correspond to interactions forming a closed loop,
and f and d do not. In particular, a1 is used for three
proteins A, B, and C that interact with each other; if
interactions having a large a1 value appear frequently in
an interaction network, the proteins participating in the
network are highly interconnected.
Implementation
Principal-component analysis was performed with R (a
language and environment for statistical computing and
graphics, http://www.r-project.org/). All other analyses
were done by means of Perl scripts we developed.
RESULTS
Assessment of new interaction generality
The original IG1 measure was based on the idea that
interactions observed in a complicated interaction net-
work are likely to be true positives. The IG1 value was
simply defined as the number of proteins that interact
with only one of the target interacting pair (Saito et al.,
2002). Interactions with low IG1 values were more likely
to be reproducible in independent assays. However, the
topological properties of the protein interaction network
beyond the target interacting pair were not considered
in the IG1. For example, IG1 values for both MED8–
MED4 and YMD8–EST1 interactions were three, even
though MED8–MED4 seems to be involved in a more
complicated interaction network and is experimentally
more certain than the YMD8–EST1 interaction (Fig. 2).
Actually, both the MED8 and MED4 proteins are known
758
by guest on July 16, 2011
bioinformatics.oxfordjournals.org
Downloaded from

Page 4

Protein–protein interaction generality measure
0
-6
0.05
0.1
0.15
0.2
0.25
-4-2
0
2
4
6
8
1012141618
20
IG2 value
Rate
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Rate(cummulative)
Rep
non-Rep
Rep
non-Rep
Fig. 3. Distributions of IG2 values for reproducible and non-reproducible interactions. The histogram shows the frequencies of interactions
falling in the specified ranges of IG2 values. The lines show the cumulative proportion of interactions. Rep and non-Rep indicate reproducible
and non-reproducible interactions, as defined in Materials and Methods.
to be transcription regulation mediators, whereas YMD8
and EST1 have different functions (chromatin chro-
mosome structure/DNA synthesis and small molecule
transport, respectively). To overcome this inaccuracy, we
developed IG2, in which principal-component analysis
of the topological properties is incorporated into the
evaluation of the reliability of the interaction. First, we
classified proteins that directly interact with the target
interacting pair A–B into one of five groups (a1, a2,
l, f and d) according to the topological properties of
its interaction network (Fig. 1 and see Materials and
Methods). Then, to convert the numbers of proteins that
belong to each class (Na1, Na2, Nl, N f , Nd) into a
single IG2 value, each of the five numbers is weighted
and summed. Principal-component analysis determines
the weight for each number, based on the correlations
between these numbers, so that a single summed value
(IG2) can represent the five numbers (Na1, Na2, Nl,
N f , Nd). By applying the computational framework
described in Materials and Methods to our interaction
data set, p1 = (−0.057, 0.0963, 0.179, 0.920, 0.331)T
and k1 = 5.603 were obtained. The IG2 value for the
MED8–MED4 interaction (−4.17) is now very different
from that for YMD8–EST1 (−0.34).
IG2 values ranging from 52.98 to –6.35 were obtained
foralltheinteractionswecollected.Toinvestigatewhether
the IG2 value may be useful for assessing the validity of
candidate protein–protein interaction pairs, distributions
of IG2 values for reproducible and non-reproducible
interactions were calculated. We expected that most
reproducible interactions should be true positives, whereas
the non-reproducible interactions should contain many
false positives. As shown in Figure 3, the IG2 values for
reproducibleinteractionsaresignificantlylowerthanthose
of non-reproducible ones, suggesting that the IG2 value
can be used to select reliable interactions (average IG2
values for reproducible and non-reproducible interactions
are −2.904 and 0.817 respectively; P < 1.37 × 10−41).
Next, we investigated the mean IG2 value for the
relatively reliable protein interaction data sets. Deane et
al. (2002) found 3003 interactions that they considered
reliable, by using information on gene expression and
paralogous proteins. The average IG2 value for our
interactions that are contained in Deane et al.’s data
set is −1.07, and the average IG2 value for those that
are not is +0.80 (P < 1.1 × 10−9), again showing
that lower IG2 values tend to occur for true protein–
protein interactions. In addition, Mering et al. showed
that interactions confirmed by more than one method,
such as by both two-hybrid and tagged proteins/mass
spectrometry (protein complex data) methods, are reliable.
The average IG2 value for protein–protein interactions
that were verified by the protein-complexing experiments
of Gavin et al. (2002) or Ho et al. (2002) is −2.48,
759
by guest on July 16, 2011
bioinformatics.oxfordjournals.org
Downloaded from

Page 5

R.Saito et al.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
00.20.40.60.81
Rate of nonreproducible below particular threshold
Rate of reproducible below particular threshold
IG2
IG1
Random
-6
-5
-4
-3
1
2
3
Fig. 4. Proportions of reproducible and non-reproducible interactions below various IG thresholds. Plots corresponding to IG thresholds from
−6 to 20 (left to right) for IG2 (diamonds) and from 1 to 21 (left to right) for IG1 (squares) are shown. Some of the plots are labeled with
corresponding IG threshold values. Theoretical proportions when the interactions are selected randomly regardless of their IG values are
plotted in triangles.
whereas the average IG2 value for interactions not verified
by their experiments is +0.39 (P < 2.72 × 10−16;
complex data were converted to binary protein–protein
interaction data by considering that the bait interacts
directly with each protein in the complex; Bader and
Hogue, 2002).
Another way to evaluate the utility of the IG2 measure
is to calculate the proportions of reproducible and non-
reproducible interactions falling below particular IG2
values. When we construct an interaction network by
selecting interactions with IG2 values below a certain
threshold, the IG2 measure is assumed to be useful if
the set of interactions below that threshold contains many
reproducible and few non-reproducible interactions. We
used this approach to compare the IG1 and IG2 measures.
Figure 4 shows the proportions of reproducible and non-
reproducible interactions below various IG2 and IG1
thresholds, among all reproducible and non-reproducible
interactions, respectively. IG2 (diamonds) performs better
than IG1 (squares) for all thresholds, except where the
proportion of reproducible interactions is close to 1.
However, this region is not important in constructing
a reliable interaction network, since the rate of non-
reproducible interactions is also close to 1 in the region.
Most of the computational time needed to calculate
IG2 is spent on calculating (a1, a2, l, f , and d) for
each interaction. Theoretically, the computational time for
calculating (a1, a2, l, f , and d) for each interaction scales
as n3, where n is the number of interacting partners for
each protein, but the computational time for calculating
IG1 is proportional to n. However in practice, proteins
having many interaction partners are rather few, which
reduces the calculation time. In fact, the real elapsed time
to calculate IG2 for 3066 interactions was <1.5 h on a
Pentium 4-based machine running Linux.
Functional associations and expressional
correlations become clear in a reliable
protein–protein interaction network
Interacting proteins generally share a common function
and a common localization (‘guilt-by-association’ princi-
ple; Oliver, 2000). Approximately 63% of interacting pro-
teins have at least one common cellular role (as defined
in the Yeast Proteome Database; Costanzo et al., 2001),
and 73–76% of them have at least one common cellular
localization (Hishigaki et al., 2001; Schwikowski et al.,
2000). We investigated the accuracy of the IG2 measure
by eliminating unreliable interactions and comparing its
performance with that of IG1.
Figure 5a and b show the proportions of interacting
protein pairs having common cellular roles and com-
mon localizations at various IG thresholds. As the IG2
threshold is decreased, the proportion of interacting pairs
with common cellular roles and localizations increases,
760
by guest on July 16, 2011
bioinformatics.oxfordjournals.org
Downloaded from