Protein interaction prediction for mouse pdz domains using dipeptide composition features

Abstract

The PDZ domain is one of the largest families of protein domains that are involved in targeting and routing specific proteins in signaling pathways. PDZ domains mediate protein-protein interactions by binding the C-terminal peptides of their target proteins. Using the dipeptide feature encoding, we develop a PDZ domain interaction predictor using a support vector machine that achieves a high accuracy rate of 82.49%. Since most of the dipeptide compositions are redundant and irrelevant, we propose a new hybrid feature selection technique to select only a subset of these compositions that are useful for interaction prediction. Our experimental results show that only approximately 25% of dipeptide features are needed and that our method increases the accuracy by 3%. The selected dipeptide features are analyzed and shown to have important roles on specificity pattern of PDZ domains.

Full text

Page 1

Protein Interaction Prediction for Mouse PDZ
Domains Using Dipeptide Composition Features
Songyot Nakariyakul∗†, Zhi-Ping Liu∗and Luonan Chen∗‡
∗Key Laboratory of Systems Biology, SIBS-Novo Nordisk Translational Research Centre for Pre-Diabetes,
Shanghai Institutes for Biological Sciences, Chinese Academy of Sciences, Shanghai 200233, China
†Department of Electrical and Computer Engineering, Thammasat University,
Khlong Luang, Pathumthani 12120, Thailand
‡Collaborative Research Center for Innovative Mathematical Modelling,
Institute of Industrial Science, University of Tokyo, Tokyo 153-8505, Japan
Email: nsongyot@engr.tu.ac.th; zpliu@sibs.ac.cn; lnchen@sibs.ac.cn
Abstract—The PDZ domain is one of the largest families
of protein domains that are involved in targeting and routing
specific proteins in signaling pathways. PDZ domains mediate
protein-protein interactions by binding the C-terminal peptides
of their target proteins. Using the dipeptide feature encoding,
we develop a PDZ domain interaction predictor using a support
vector machine that achieves a high accuracy rate of 82.49%.
Since most of the dipeptide compositions are redundant and
irrelevant, we propose a new hybrid feature selection technique
to select only a subset of these compositions that are useful for
interaction prediction. Our experimental results show that only
approximately 25% of dipeptide features are needed and that our
method increases the accuracy by 3%. The selected dipeptide
features are analyzed and shown to have important roles on
specificity pattern of PDZ domains.
Index Terms—Dipeptide compositions; feature selection; PDZ
domain; protein interaction.
I. INTRODUCTION
The PDZ (PSD-95/Discs-large/ZO-1) domain family is an
important signaling protein that is involved in the development
of multi-cellular organisms [1]. Many PDZ domains are key
components in maintaining cell polarity, facilitating intercel-
lular signaling system, and regulating synaptic development
[2], [3]. They are composed of approximately 80 to 90 amino
acid residues folded into six β strands (β1-β6) and two α
helices (α1, α2). Prior studies showed that PDZ domains
selectively bound C-terminal peptide sequences from voltage
gated potassium channels and N-methyl-D-aspartate receptors
[4], [5], specifically on residues up to -8 position of the peptide
ligand (last residue numbered zero) [6]. Furthermore, many
PDZ domains display promiscuity and bind to more than
one ligand. However, experimental methods to determine the
interaction specificity of the PDZ domains are time-consuming
and expensive. Thus, a computational method that can provide
accurate prediction is highly demanded.
Several computational methods have been proposed to pre-
dict interaction specificity of PDZ domains. Chen et al. [7]
proposed an extension of a position-specific scoring matrix
that predicted interactions between the 82 mouse PDZ domains
and 93 peptides based on their primary sequences. They
reported an area under the receiver operating characteristic
(AUC) value of 0.87. Eo et al. [8] used amino acid contact
matrices and physicochemical distance matrix to encode the
protein complex into a feature vector. A support vector ma-
chine (SVM) classifier was employed to identify G protein-
coupled receptors-binding PDZ domain proteins. Recently,
Kalyoncu et al. [9] used trigram amino acid frequencies for
feature encoding and a random forest classifier to build a
model to predict the binding interactions of PDZ domains
and peptide sequences. Resampling was used to address the
problem of the imbalanced data set. They obtained an accuracy
of 79.8% on the validation set of 27 binding and 62 non-
binding interactions.
In this work, we propose to use dipeptide compositions as
feature encoding to predict PDZ domain-peptide interactions
and employing an SVM classifier to build our predictor.
Dipeptide compositions have been shown to give useful infor-
mation in prior protein-related work [10]–[12]. We compare
our method with other feature encoding techniques based
on primary sequences. Our experimental results demonstrate
that our predictor can obtain a high prediction performance
(accuracy of 82.49% and AUC of 0.8920). To further improve
the prediction results, we develop a new hybrid feature se-
lection algorithm named the mRMR BIRS algorithm that is
a combination of the minimal-redundancy-maximal-relevance
(mRMR) algorithm [13] and the best incremental ranked
subset (BIRS) algorithm [14]. We find that approximately
25% of dipeptide features are needed for interaction prediction
and that our proposed method increases the accuracy by 3%.
Analysis of selected dipeptide compositions is also given.
II. MATERIALS AND METHODS
A. Data Set
We used the PDZ interaction data set provided in [9]. The
data set was originally retrieved from the study of Stiffer
et al. [15], which contained interaction data of 85 mouse
PDZ domains and 181 mouse peptides. There are a total of
731 binding and 1361 non-binding interactions available for
testing. The data set is imbalanced due to its nature. These
interaction data were confirmed by fluorescence polarization
experiments. The last 10 residues (up to -9 position) of each
2011 IEEE International Conference on Systems Biology (ISB)
978-1-4577-1666-9/11/$26.00 ©2011 IEEE
129
Zhuhai, China, September 2–4, 2011

Page 2

peptide ligand are considered in our computational method due
to the specificity of the PDZ domains. For more information
about the data set, see [9].
B. Feature Encoding
The 400 dipeptide compositions of each protein sequence
are computed using the following expression (1)
Compdipeptide(i,j) =
nij
L − 1,1 ≤ i,j ≤ 20
(1)
where i, j stand for the distribution of amino acid i followed
by amino acid j, nijis the number of residues of amino acid
i followed by amino acid j, and L is the total number of
residues in the protein sequence. For each binding/non-binding
interaction, a PDZ domain and a peptide ligand are encoded
into two vectors, each with 400 dipeptide compositions. We
then concatenate two vectors into an 800-feature vector to
represent each interaction.
C. Feature Selection
Feature selection refers to search algorithms that select a
subset of features from an initial set of n features, where a
criterion function J is used to evaluate the quality of each
candidate subset. It is mainly used for identifying important
features and improving classification results. Depending on
the criterion function J used, feature selection methods can
be categorized as filter or wrapper. Filter methods rely on the
intrinsic properties of the data such as distance, dependency,
and consistency and select subsets without any knowledge of
the learning algorithm. Wrapper methods use the performance
of a predetermined learning algorithm as the criterion function
to select a subset. The wrapper method generally achieves
better performance than the filter method, but it is also more
computationally expensive. Since there are a total of 800
dipeptide features (400 for the PDZ domains and another 400
for the peptide ligands) for our work, we are interested in a
wrapper method that is highly effective and computationally
efficient.
In this work, a modified version of the best incremental
ranked subset (BIRS) algorithm for feature selection is pro-
posed. We first discuss the original BIRS algorithm [14] and
then present its modification. The BIRS algorithm is a wrapper
method that contains two phases; in the first phase, all of the
n features in the set are ranked according to some evaluation
measure. In the second phase, the search proceeds from the
best to the worst ranked feature, and a feature is selected if
adding it to the currently selected feature subset improves the
accuracy significantly. That is, the algorithm starts by selecting
the best ranked feature from the list. It then considers adding
the second best ranked feature to the best one if and only if the
resultant subset increases the accuracy rate significantly. If the
accuracy obtained by adding the second best ranked feature to
the set is not significantly better, the feature is discarded, and
the third best ranked feature is considered next, and so on.
A Student’s paired two-tailed t-test is conducted to determine
the statistical significance degree of difference between the
accuracies of each subset using a fivefold cross-validation. The
algorithm terminates when it reaches the worst ranked feature.
Thus, BIRS runs in linear time and selects only relevant and
irredundant features.
In the original BIRS work [14], the authors ordered the
features according to their individual accuracy rates (the per-
formance of a pre-defined classifier built with a single feature).
Since ranking of all features in the first stage plays an impor-
tant role on the performance of the algorithm, we thus propose
using the minimal-redundancy-maximal-relevance (mRMR)
algorithm [13] to order the feature set. The mRMR algorithm
is a well-known filter search technique that selects feature sub-
sets based mutual information. It is fast and shown to perform
well in many applications. We thus expect it to give a better
list of ranked features than that ranked by individual accuracy
rates as done in prior work [14]. Moreover, to measure the
significance of adding a feature to the current subset, we
compute the difference between the AUC values of each subset
rather than the accuracy rates, since our data set is imbalanced.
We name this modified algorithm the mRMR BIRS algorithm.
D. Support Vector Machines and Performance Evaluation
We choose the SVM classifier with a radial basis function to
perform the classification, since it provides high classification
results and is fairly resistant to feature selection. The software
LIBSVM [16] version 3.0 is employed in this work. The
regularization parameter C and kernel parameter γ in the
SVM are selected by using a grid search approach. The SVM
classifier is trained by using a fivefold cross-validation to
maximize an area under the receiver operating characteris-
tic (AUC), since our data set is imbalanced. The receiver
operating characteristic (ROC) is a plot of the true positive
rate (TPR) versus false positive rate (FPR). We also provide
the accuracy (ACC) rate to measure the performance of our
method. TPR, FPR, and ACC are defined as follows.
TPR =
TP
TP + FN
FP
FP + TN
TP + TN
(2)
FPR =
(3)
ACC =
TP + TN + FP + FN
(4)
where true positive (TP) is the number of binding interactions
correctly classified. True negative (TN) is the number of non-
binding interactions correctly classified. False positive (FP) is
the number non-binding interactions misclassified as binding
interactions. False negative (FN) is the number of binding
interactions misclassified as non-binding interactions.
III. RESULTS AND DISCUSSION
A. Feature Encoding Comparisons
We now compare our dipeptide composition model with
other feature encoding proposed for predicting protein-protein
interactions in the literature [11], [17], [18]. Amino acid
composition (AAC) has been used in many prior protein-
related work [11]. It is defined as the frequencies of 20 amino
acids in a protein sequence. In the triad frequency model
2011 IEEE International Conference on Systems Biology (ISB)
978-1-4577-1666-9/11/$26.00 ©2011 IEEE
130
Zhuhai, China, September 2–4, 2011

Page 3

[17], the 20 amino acids are grouped into 7 different classes
according to their dipoles and volumes of the side chains.
These classes contain [AGV], [ILFP], [YMTS], [HNQW],
[RK], [DE], and [C] amino acids, respectively. Frequencies
of three consecutive classes in each protein sequence are
then computed and used as features. Pseudo amino acid
composition (PseAAC) [18] incorporates both sequence-order
information and protein properties to represent each protein
sequence. The first 20 features of the PseAAC contain the
AAC information, and the additional λ features represent the
sequence-order information calculated by the hydrophobicity
value, hydrophilicity value, and side-chain mass. We choose
λ = 5, since this gives the highest AUC value.
Fig. 1 shows the ROC curves of the four feature encoding
models. As seen, the AAC model gives the worst performance
among the four models in many regions. This is expected,
since it does not utilize the sequence-order information. The
PseAAC model is performing slightly better than the triad
model, but a clear difficulty with the PseAAC model lies
in interpreting and understanding the model. The proposed
dipeptide model is shown to outperform the other models
in most regions. Table I summarizes the results of the four
feature encoding models using a fivefold cross-validation test.
Although the FPR of the PseAAC model (12.28%) is slightly
lower than that of the dipeptide model (12.87%), its TPR
(69.45%) is the worst one among the four algorithms. As
seen in Table I, the ACC rate (82.49%) and AUC (0.8920) of
the dipeptide model are highest. However, the TPR (73.84%)
obtained using the dipeptide model is somewhat low due to the
imbalance of the data set. We expect to use feature selection
to combat this problem and improve the prediction results.
0 0.20.4
False positive rate
0.6 0.81
0
0.2
0.4
0.6
0.8
1
True positive rate


AAC
Triad
PseAAC
Dipeptide
Fig. 1. ROC curves of four different feature encoding models.
B. Feature Selection Results
Since the lengths of PDZ domains and peptides are short,
many encoded dipeptide features contains zeros. These fea-
TABLE I
FIVEFOLD CROSS-VALIDATION PREDICTION RESULTS FOR INTERACTION
PREDICTION OF PDZ DOMAINS USING DIFFERENT FEATURE ENCODING
MODELS.
Model
AAC
Triad
PseAAC
Dipeptide
No. of features
40
686
50
800
TPR (%)
71.64
71.78
69.45
73.84
FPR (%)
15.15
13.90
12.28
12.87
ACC (%)
80.24
81.10
81.34
82.49
AUC
0.8726
0.8748
0.8837
0.8920
tures may not contribute to prediction results and can be
deemed irrelevant. We propose the mRMR BIRS algorithm
to select only relevant and irredundant dipeptide features for
interaction prediction. To determine the statistical significance
degree of difference between the AUC values of each subset in
the mRMR BIRS algorithm, the confidence level is chosen to
be p < 0.5 due to the small sample size. We thus expect many
features to be selected by our algorithm. For comparison, we
also apply the original BIRS feature selection algorithm to
select dipeptide features.
Table II shows the number of selected dipeptide features and
the prediction performances of the original BIRS algorithm
and our mRMR BIRS algorithm. As seen, the BIRS algorithm
selects only 54 dipeptide features and yields poorer results
than using all 800 dipeptide features (see Table I). The reason
is that the BIRS algorithm does not employ a good feature
ranking. Our proposed mRMR BIRS algorithm, on the other
hand, uses the mRMR method to provide an initial ranking,
which is shown to be very effective. The prediction results of
the mRMR BIRS algorithm are much better than those of the
BIRS algorithm and those using all 800 dipeptides; the ACC
rate increases from 82.49% to 85.17%, and the AUC value
increases from 0.8920 to 0.9110 using our feature selection
method. The mRMR BIRS algorithm selects 215 dipeptide
features (approximately 25% of the original 800 features),
which shows that many dipeptide features are redundant and
irrelevant. Out of 215 features selected by mRMR BIRS, 102
features are selected from PDZ domains and the other 113
features are chosen from peptides. In terms of computational
complexities, the mRMR BIRS algorithm takes only 10 min-
utes to perform the search, while the BIRS algorithm needs
more than 20 minutes. Thus, our mRMR BIRS algorithm is
faster and more effective.
TABLE II
FIVEFOLD CROSS-VALIDATION PREDICTION RESULTS FOR INTERACTION
PREDICTION OF PDZ DOMAINS USING DIPEPTIDE FEATURES AFTER
FEATURE SELECTION.
Model
BIRS
mRMR BIRS
No. of features
54
215
TPR (%)
61.10
76.85
FPR (%)
11.40
10.37
ACC (%)
79.00
85.17
AUC
0.8288
0.9110
We now analyze some important dipeptide features selected
by our feature selection algorithm. For example, the most
important (best ranked) dipeptide selected by our algorithm is
‘Glu-Thr’ of the peptide ligand, which is supported by prior
finding [6] that many PDZ domains such as those of the Discs
2011 IEEE International Conference on Systems Biology (ISB)
978-1-4577-1666-9/11/$26.00 ©2011 IEEE
131
Zhuhai, China, September 2–4, 2011

Page 4

Large Protein bind to the C-terminal motifs of the peptide lig-
and with the sequence of Glu-(Ser/Thr)-Xxx-(Val/Ile)-COOH,
where Xxx represents any amino acid. Many ‘Ser-Xxx’ motifs
such as ‘Ser-Leu’, ‘Ser-Asn’, ‘Ser-Gln’, and ‘Ser-Ser’ of the
peptides are also selected by our mRMR BIRS algorithm. ‘Ile-
Arg’ of the PDZ domain chosen by our method is also found
to play an important role in forming hydrophobic contact of
the first PDZ domain of NHREF1 with amino acid Leu of the
peptide ligand [19]. Furthermore, Bezprozvanny and Maximov
[20] reported that eight PDZ domains in hINADL-5 bound
to neurexin Ia, whose terminus is ‘Glu-Tyr-Tyr-Val’. This is
also in agreement with our prediction model that selects ‘Tyr-
Val’ of the peptide as an important dipeptide motif. Thus, our
selected dipeptides can be used as a guide for future study of
prediction interaction specificity of PDZ domains.
IV. CONCLUSIONS
This study of PDZ domain-peptide interactions has two
aims. First, we compared the prediction performance of the
dipeptide composition model to those of three other feature
encoding models based on primary sequences. We found
that the SVM-based predictor based on the dipeptide model
successfully achieved the fivefold cross-validation accuracy of
82.49%, which is slightly higher than those obtained using
the other models. Second, we proposed a new mRMR BIRS
feature selection algorithm to further improve the prediction
results and to identify important dipeptide motifs. The pro-
posed method was shown to outperform the original BIRS
algorithm and increased the prediction accuracy from 82.49%
to 85.17%. Many important motifs of PDZ domains and
peptides were identified by our method. These encouraging
results could be used to facilitate future study on PDZ domain
interactions. As a future topic, we will further consider to
employ the network-based technique to improve the accuracy
of the prediction [21], [22].
ACKNOWLEDGMENTS
The authors would like to thank S. Kalyoncu,O. Keskin, and
A. Gursoy for their helpful suggestions and for providing the
data set used in this work. S. Nakariyakul was supported by the
CAS Fellowship for Young International Scientist with Grant
No. 2010Y1SB10 and NSFC with Grant No. 31050110435.
This work was also supported by the Chief Scientist Program
of SIBS of CAS with Grant No. 2009CSP002 (L. Chen)
and by the Knowledge Innovation Program of SIBS of CAS
with Grant No. 2011KIP203 (Z.P. Liu), and supported by
NSFC under Grants No. 61072149 and No. 91029301 (L.
Chen and Z.P. Liu), the Knowledge Innovation Program of
CAS with Grant No. KSCX2-EW-R-01 (L. Chen and Z.P.
Liu), and supported by the Key Project of Shanghai Education
Committee (B.10-0412-08-001),Japan (JSPS) FIRST Program
(L. Chen) and Shanghai NSF under Grant No. 11ZR1443100
(Z.P. Liu).
REFERENCES
[1] C.P. Ponting, “Evidence for PDZ domains in bacteria, yeast and plants,”
Protein Sci 6, pp. 464-468, 1997.
[2] T. Pawson, and P. Nash, “Assembly of cell regulatory systems through
protein interaction domains,” Science 300, pp. 445-452, 2003.
[3] K.K. Dev, “Making protein interactions druggable: Targeting PDZ do-
mains,” Nat Rev Drug Discov 3(12), pp. 1047-1056, 2004.
[4] H. Kornau, R.L. Schenke, M. Kennedy, and P. Seeburg, “Domain inter-
action between NMDA receptor subunits and the postsynaptic density
protein PSD-95,” Science 269, pp. 1737-1740, 1995.
[5] J. Schultz, U. Hoffmuller, G. Krause, J. Ashurst, M.J. Macias, P.
Schmieder, J. Schneider-Mergener, and H. Oschkinat, “Specific inter-
actions between the syntrophin PDZ domain and voltage-gated sodium
channels,” Nat Struct Biol. 5, pp. 19-24, 1998.
[6] Z. Songyang, A.S. Fanning, C. Fu, J. Xu, S.M. Marfatia, A.H. Chishi,
A. Crompton, A.C. Chan, J.M. Anderson, L.C. Cantley, “Recognition of
unique carboxyl-terminal motifs by distinct PDZ domains,” Science 275,
pp. 73-77, 1997.
[7] J.R. Chen, B.H. Chang, J.E. Allen, M.A. Stiffler, and G. MacBeath,
“Predicting PDZ domain-peptide interactions from primary sequences,”
Nat Biotechnol 26(9), pp.1041-1045, 2008.
[8] H.S. Eo, S. Kim, H. Koo, and W. Kim, “A machine learning based method
for the prediction of G protein-coupled receptor-binding PDZ domain
proteins,” Mol Cells 27, pp. 629-634, 2009.
[9] S. Kalyoncu, O. Keskin, and A. Gursoy, “Interaction prediction and
classification of PDZ domains,” BMC Bioinformatics 11, 357, 2010.
[10] Q.B. Gao, Z.C. Jin, X.F. Ye, C. Wu, J. Lu, and J. He, “Improving the
classification of nuclear receptors with feature selection,” Protein Pept
Lett 16(7), pp. 823-829, 2009.
[11] M. Rashid, S. Ramasamy, and G.P. Raghava, “A simple approach for
predicting protein-protein interactions,” Curr Protein Pept Sci 11(7), pp.
589-600, 2010.
[12] S. Nakariyakul, Z.-P. Liu, and L. Chen, “Detecting thermophilic proteins
through selecting amino acid and dipeptide composition features,” Amino
Acids, doi:10.1007/s00726-011-0923-1, 2011.
[13] H.C. Peng, F.H. Long, and C. Ding, “Feature selection based on
mutual information: criteria of max-dependency, max-relevance, and min-
redundancy,” IEEE Trans Pattern Anal Mach Intel 27, pp. 1226-1238,
2005.
[14] R. Ruiz, J.C. Riquelme, and J.S. Aguilar-Ruiz, “Incremental wrapper-
based gene selection from microarray data for cancer classification,”
Pattern Recog 39, pp. 2383-2392, 2006.
[15] M.A. Stiffler, J.R. Chen, V.P. Grantcharova, Y. Lei, D. Fuchs, J.E. Allen,
L.A. Zaslavskaia, and G. MacBeath, “PDZ domain binding selectivity is
optimized across the mouse proteome,” Science 317, pp. 364-369, 2007.
[16] C.-C. Chang, and C.-J. Lin, “LIBSVM: a library for support vector
machines,” ACM Trans Intell Syst Technol 2, pp. 27:1-27:27, 2011.
[17] J. Shen, J. Zhang, X. Luo, W. Zhu, K. Yu, K. Chen, Y. Li, and H.
Jiang, “Predicting protein-protein interactions based only on sequence
information,” Proc Natl Acad Sci USA 104(11), pp. 4337-4341, 2007.
[18] L. Liu, Y. Cai, W. Lu, K. Feng, C. Peng, and B. Nu, “Prediction
of protein-protein interactions based on PseAA composition and hybrid
feature selection,” Biochem Biophys Res Commun 380, pp. 318-322,
2009.
[19] S. Karthikeyan, T. Leung, and J.A. Ladias, “Structural basis of the
Na+/H+ exchange regulatory factor PDZ1 interaction with the carboxyl-
terminal region of the cystic fibrosis transmembrane conductance regula-
tor,” J Biol Chem 276(23), pp. 19683-19686, 2001.
[20] I. Bezprozvanny, and A. Maximov, “Classification of PDZ domains,”
FEBS Lett 509, pp. 457-462, 2001.
[21] L. Chen, R. Wang, and X. Zhang, Biomolecular Network: Methods and
Applications in Systems Biology, Wiley, London, 2009.
[22] L. Chen, R. Wang, C. Li, and K. Aihara, Modelling Biomolecular
Networks in Cells: Structures and Dynamics, Springer, Berlin, 2010.
2011 IEEE International Conference on Systems Biology (ISB)
978-1-4577-1666-9/11/$26.00 ©2011 IEEE
132
Zhuhai, China, September 2–4, 2011