Engineering, 2013, 5, 409-412 Published Online October 2013 (
Copyright © 2013 SciRes. ENG
Prediction Method of Protein Disulfide Bond Based on
Pattern Sel ection*
Pengfei Sun, Yuanq uan Cui#, Tiankai Chen, Ying Zhao
College of Computer Science and Technology Harbin Normal University, Harbin, C hina
Email: sunpe, 1512428681
Received 2013
The effect of the different training samples is different for the classifier when pattern recognition system is established.
The training samples were selected randomly in the past protein disulfide bond prediction methods, therefore the pre-
diction accuracy of protein contact was reduced. In order to improve the influence of training samples, a prediction me-
thod of protein disulfide bond on the basis of pattern selection and Radical Basis Function neural network has been
brought forward in this paper. The attributes related with protein disulfide bond are extracted and coded in the method
and pattern selection is used to select training samples from coded samples in order to improve the precision of protein
disulfide bond prediction. 200 proteins with disulfide bond structure from the PDB database are encoded according to
the encoding approach and are taken as models of training samples. Then samples are taken on the pattern selection
based on the nearest neighbor algorithm and corresponding prediction models are set by using RBF neural network. The
simulation experiment result indicates that this method of pattern selection can improve the pre diction accuracy of pro-
tein disulfide bond.
Keywords: Protein Disulfide Bond; Neural Network; Nearest Neighbor Algorithm; Pattern Selection
1. Introduction
The protein disulfide bond is important component for
many proteins; it can maintain the stability and function
activity of proteins. The correct orientation of protein di-
sulfide bond is very important to grasp the relationship of
the protein structure and its biological function. There-
fore, the solution to predict protein disulfide bond has
great significance to predict the protein spatial structure
and the protein function, but it is still hard to pr edict pro-
tein disulfide bond [1].
Some methods of predicting contact have been devel-
oped to solve the problem. Currently, there are many
methods spreading internationally, such as artificial neural
network, SVM, Genetic Programming, Hidden Markov
Model and so on [2-4]. However, generally speaking, the
predicting precision of these methods is not high enough.
In order to enhance the predicting precision of the protein
disulfide bond, a prediction method of protein disulfide
bond on the basis of pattern selection and RBF neural
network have been brought forward in this paper. In the
method proteins with disulfide bond structure from the
PDB database are encoded according to the encoding
approach and are taken as models of training samples.
Then samples are taken on the pattern selection based on
the nearest neighbor algorithm and corresponding predic-
tion models are set by using RBF neural network. As a
result, the experiment indicates that the method could en-
hance the predicting accuracy of the disulfide bond ef-
2. Definition of K-NN Algorithm
In pattern recognition, the k-nearest neighbor algorithm
(k-NN) is a method for classifying objects based on clos-
est training examples in the feature space. k-NN is a ty pe
of instance-based learning, or lazy learning where the
function is only approximated locally and all computa-
tion is deferred until classification; the k-nearest neigh-
bor algorithm is amongst the simplest of all machine
learning algorithms [5].
K-nearest neighbor algorithm(k-NN) is very simple,
{, ,}
be N, n-dimensional design sam-
ples, it is required to compute the k-nearest neighbors of
a test sample X among
{, ,}
, as measured
by an appropriate distance function d. The training exam-
ples are vectors in a multidimensional feature space, each
with a class label. The training phase of the algorithm
consists only of storing the feature vectors and class la-
bels of the training samples. I n the classification ph ase, k
*The project was s upported b y Scient ific Resea rch Fun
d of Heilong jiang
Provincial Education Department under Grant No. 11551128.
#Corresponding author.
Copyright © 2013 SciRes. ENG
is a user-defined constant, and an unlabelled vector (a
query or test point) is classified by assigning the label
which is most frequent among the k training samples
nearest to that query point.
3. Selection of Protein Nature and Encoding
In the natural world protein has many and varied attributes.
It is not realistic to make all the attributes determine the
methods as the condition of forming the disulfide bond.
In this research, based on the former results, several kinds
of important attributes are selected to serve as the input
items of the predicting models, and encode these data ac-
cording to their characteristics.
Attribute 1: hydrophobic of amino acid
In the water medium, globular protein folding always
favors in burying the hydrophobic amino acid to the in-
side of the protein. This phenomenon is called hydro-
phobic of amino acid. It holds the prominent status in
stabling the protein three-dimensional structure. The for-
mer researching results indicate that there are differences
between the disulfide bond and non disulfide bond in the
existence ratio of hydrophobic amino acid. The results
reveal that the hydrophobic of amino acid is very impor-
tant to the forming process of the disulfide bond. There-
fore, the property of amino acid can be used as input of
prediction model. According to hydrophobic of amino
acid, hydrophobic amino acid is coded as 1 and non hy-
drophobic amino acid is encoded as 0.
Attribute 2: protein secondary structure [6 ]
Protein secondary structure includes protein folding
information which is significant to predict and recon-
struct the protein 3D structure. Simultaneously, it is im-
portant to the prediction of the protein conjunction which
is proceeding in this thesis. The protein secondary struc-
ture from the data in the research is chosen from second-
ary structure database DSSP. DSSP is a database of sec-
ondary structure assignments for all proteins in the Pro-
tein Data Bank. To any protein in the database of protein
3D structure PDB, the corresponding secondary structure
can be derived by its three-dimensional structure. Ac-
cording to the protein secondary structure, the structure
code of α is 00, the structure code of βis 01, and other
structure cod e is 10.
Attribute 3: evolution information of protein [7]
In the research of predicting the protein secondary
structure, it can be found that using the evolution infor-
mation of protein will increase the predicting accuracy
obviously. It reveals that the evolution information of
protein contains the important information of protein
structure formation. Thus this research introduces evolu-
tion information of protein to predict the protein disulfide
bond. In this research, the protein evolution information
gains from the HSSP database. The HSSP is a database
of protein secondary structures derived by aligning to
each protein of known structure all sequences deemed
homologous. HSSP contains the sequence information
which is based on the sequences ratio between the pro-
tein and its homology in the protein database. According
to the protein secondary structure, every amino acid is
encoded as Pi, in every position, Pi is the probability of
the presenting some kind of amino acid. The value scope
of i is 1 to 20.
Because of the consideration of the interaction be-
tween the neighboring amino acids, the sliding windows
with the length as 7 are chosen as a unit. According to
the above encoding method, the predicting amino acid
pair (i, j) is encoded and a group of 322-dimensional data
are gotten as an input vector
(, )Tij
4. Sample Selection Method Based on the
K-Nearest Neighbor Algorithm
Artificial neural network (BP), a theore tical classification
model, is raised from the simulation of the brain infor-
mation processing and the learning procedures. It puts
forward on the basis of the human science information
processing research, contains modern Neurobiology and
cognitive science, and has very strong adaptivity and
self-learning ability, and the nonlinear mapping ability,
robustness and fault-tolerant capability, etc. In recent
years, with the development of artificial neural network,
it uses in every field of bioinformatics successfully, and
the technique of artificial neural network is increasingly
becoming an important tool for solving the problems on
sequence analysis and pattern recognition of machine
learning technology. For neural network, it depends on
the quality of the training samples, it may not contain
enough information when the performance of the training
samples is too small; and if the training samples are too
large, it may be too large and make the sample redundan-
cy, increase the training time and is likely to cause over-
fitting. So the selection process of the training samples
has an important meaning on prediction modeling, and
how to choose the training sample will be the key to im-
prove the performance of classification. Through the ana-
lysis of the working principle of neural network, it is
known that neural network is one of the optimization of
the nearest neighbor classifier essentially, only its tem-
plate stored in the network structure by form of weight
values, through repeated adjusting relevance weights to
the purpose of fitting with the template. And K-Nearest
Neighbor Algorithm make representative samples as a
template directly, determine the category of the sample
according to the distance to the template. So there is no
need to iterative process of iterative adjustment. Both
compare, neural network training is complex but classi-
fication accuracy is higher, and K-Nearest Neighbor Al-
gorithm is low precision but simple and quick. Therefore
Copyright © 2013 SciRes. ENG
we can use the advantage of K-Nearest Neighbor Algo-
rithm which is simply and quickly to structure classifier
to choose the sample, and then use these subset samples
of this operation as the training sample to build a neural
network classifier. So it still maintained the high preci-
sion characteristics of neural network classifier.
In the K-Nearest Neighbor Algorithm, the boundary
samples play an important role for classification, and all
kinds of samples that near the centre are less effective,
choosing sample with K-Nearest Neighbor Algorithm is
that, trying to delete all kinds of near centre sample and
keep boundary samples, only to reduce sample size and
improve the accuracy of the classifier. However, in sam-
ple selection process, the sequence of subset used for
sample selection is a certain order, so that the samples in
front can be reserved probability, which will cause boun-
dary samples behind deleted and the class center samples
in front of the sequence retained, so the representative of
selected samples is unsatisfactory and redundancy, and
cannot reach the quality for corresponding sample selec-
tion purpose. In order to solve the problem, this study
uses an improvement K-Nearest Neighbor Algorithm to
improve the quality of selected samples. The basic work-
ing principle of this algorithm is generating a sample
subset D on the basis of the original samples set T, so as
to make sure the set T can still be correctly classified
with the condition of decreasing of samples in D. When a
sample in T could not be classified correctly, it will be
added to the sample set D, until a certain cycle ended and
the samples subset D does not change. This algorithm
through the cycle repeatedly perform the most neighbor
algorithm, in o rder to improve the representation of choos-
ing the sample can reduce redundant sample, to ensure
that all the boundary samples in the samples set T were
chosen to put in selected samples set D. The text below is
the algorithm description [8]:
Algorithm input:
(1)The initial sample set T of the training samples, se-
lected sample set
D= ∅
(2) Repeated times n of the sample choosing proce-
Algorithm output:
The samples set D contains selected samples
Algorithm process:
(1) Choose any one
from the samples set T. Store
it into the set D:
TTx= −
(2) For all samples in the set, execute the following
operation: choose any one x from T, execute nearest
neighbor search operation on x in the subset D, find the
samp l e s which is nearest from x,
minDistance( ,)
sD i
judge the classes of sample, if
()()Classxclass s
, then
DDx= ∪
TTx= −
(3) Repeated execute (2) operation n times;
Algorithm end
5. Establish men t of the Predicting Model
The artificial neural networks technology, as a kind of
important tool of machinery learning technologies, is
becoming more and more significant in bioinfo rmatics to
solve the problem of sequential analysis and pattern rec-
ognition. However, in the training process, the revision
of the all network weights and threshold is needed. There-
fore, the speed of studying is quite slow. RBF neural
network structure is a kind of network based on partial
approaches. To each training samples, it only needs to
revise few weights and threshold, thus the speed of stud-
ying is faster. Con sequently, the using of the RBF neural
network or not is the classifier to determine whether the
contact structure form.
RBF neural network [9] is a kind of back-propagation
network, and has two network levels: the hidden layer is
radial basis function layer, the output is linear layer. As
Figure 1 shows the network has Q group of input vector,
the element in every group is R, there are S1 RBF neuron
in the intermediate level, S2 linear neuron in the output
level. The output network is:
|||| *nIW Pb=−
221 2
()apurelin LWab= +
Radbas( ) is the radial basis function, generally is the
Gauss function, and purelin( ) is the linear output func-
tion. The radial basis function network simulates the ad-
justment of middle part of human brain and the neural
network structures which cover the receive territory mu-
tually. Therefore, it is a kind of network based on partial
approaches, besides that, its hidden strata node has the
mutually independent center and the width, thus has high-
er classified precision. The simulation experiment indi-
cates that comparing with BP neural network; RBF neur-
al network has the characteristics of faster speed of net-
work training and higher classified precision.
6. Result and Analysis of the Experimental
The experiment protein data used in the experiment is got
from PDB, which is a protein structure database; extract-
ing the sequence of amino acids and atomic coordinate
information of protein for the as input information of the
Figure 1. RBF Network Structure.
Copyright © 2013 SciRes. ENG
prediction model. In orig inal PDB data rate file, it usual-
ly uses SSBOND records to show that disulfide bonds
form information in peptide chain. It has two kinds of
bonds, intra-chain and inter-chain disulfide bonds, as a
result the data only account intra-chain disulfide bonds.
Process the database file from the PDB, remove the pro-
tein which contains nonstandard amino acid residues, se-
quence information deletion, or inaccurate SSBOND
sequence information. At last randomly selected 200 pro-
teins as the training sample, and select 50 pro teins as the
test sample.
The study uses the formula below as the prediction
accuracy evaluation criterion to compare the efficiency
of the prediction algorithm:
cp cp
stands for the correct count of the prediction of
disulfide bond structure,
is the amount of all. The
result is prediction accuracy.
The study calculated accuracy of the prediction of the
disulfide bond structure raised in this study, according to
this evaluation criterion. And it was compared with other
algorithms without K-NN; through the prediction result
the prediction algorithm of disulfide bond based on sam-
ple selection technology which put forward in the re-
search can improve the pre c ision o f pr edicti on (Table 1).
7. Conclusion
The study presents the protein disulfide bond structure
prediction algorithm based on sample selection technol-
ogy; it improved the selection of th e training sa mples and
classifier performance. Experiment results show that the
algorithm can improve the efficiency of the prediction
accuracy of disulfide bond structure. Test results show
that there are some wrong forecasts, in which non-disul-
fide bond structure is predicted to be disulfide bond
Table 1. Comparison of prediction accuracy.
Method RBF without K-NN RBF with K-NN
Accuracy 82.6% 83.5%
structure. So it affects the prediction accuracy. In the
future we will study how to use the structure information
of proteins to reduce the error prediction rate of non-dis-
ulfide bond structure effectively, so as to increase preci-
sion of the disulfide bond structure prediction.
[1] S. M. Muskal, R. S . Holbrook and S. H. Kim, “Prediction
of the Disulfide-Bonding State of Cysteine in Proteins,”
Protein Engineering, Vol. 3, 1990, pp. 667-672.
[2] P. Fariselli, P. Riccobelli and R. Casadio, Role of evolu-
tionary Information in Predicting the Disulfide-Bonding
State of Cysteine in Proteins,” Proteins, Vol. 36, 1999, pp.
[3] M. H. Mucchielli-Giorgi, S. Hazout and P. Tuffery, “Pre-
dicting the Disulfide Bonding State of Cysteines Using
Protein Descriptors,” Proteins, Vol. 46, 2002, pp. 243-
[4] X.-H. Shi and Y. Wang, Prediction of Disulfide Bonding
in Protein Structure Based on Method of Graph Mat ching,”
Computer Engineering and Applications, Vol. 43, 2007,
pp. 30-32.
[5] P. E. Hart, The Condensed Nearest Neighbour Rule,”
IEEE Transactions on Information Theory, Vol. 14, 1968,
pp. 515-516.
[6] J.-M. Chandonia and M. Karplus, “New Methods for
Accurate Prediction of Protein Secondary Structure,”
Proteins, Vol. 35, 1999, pp. 293-306.
[7] C. Dodge, R. Schneider and C. Sander, “The HSSP Data-
base of Protein Structure-Sequence Alignments and Fam-
ily Profiles,” Nucleic Acids Research, Vol. 26, 1998, pp.
[8] H.-W. Hao and R.-R. Jiang, “Training Sample Selection
Method for Neural Networks Based on Nearest Neighbor
Rule,” Acta Automatica Sinica, Vol. 33, 2007, pp. 1247-
[9] W. Li and Y. Hori, “An Algorithm for Extracting Fuzzy
Rules Based on RBF Neural Network,” IEEE Transac-
tions on Industrial Electronics, Vol. 53, No. 4, 2006.