70 Protein & Peptide Letters, 2012, 19, 70-78

A Novel Sequence-Based Method for Phosphorylation Site Prediction with

Feature Selection and Analysis

Zhi-Song He1,*, Xiao-He Shi2, Xiang-Ying Kong2,4, Yu-Bei Zhu3 and Kuo-Chen Chou5

1
CAS-MPG Partner Institute for Computational Biology, Shanghai Institutes for Biological Sciences (SIBS), Chinese
Academy of Sciences (CAS), 320 Yueyang Road, Shanghai 200031, China; 2Institute of Health Sciences, SIBS, CAS and
Shanghai Jiaotong University School of Medicine (SJTUSM), China; 3Department of Chemistry, College of Sciences,
Shanghai University, 99 Shang-Da Road Shanghai 200444, China; 4State Key Laboratory of Medical Genomics, Ruijin
Hospital, Shanghai Jiaotong University, 197 Rui Jin Road II, Shanghai 200025, People’s Republic of China; 5Gordon Life
Science Institute, 13784 Torrey Del Mar, San Diego, California 92130, USA

Abstract: Phosphorylation is one of the most important post-translational modifications, and the identification of protein
phosphorylation sites is particularly important for studying disease diagnosis. However, experimental detection of phos-
phorylation sites is labor intensive. It would be beneficial if computational methods are available to provide an extra ref-
erence for the phosphorylation sites. Here we developed a novel sequence-based method for serine, threonine, and tyro-
sine phosphorylation site prediction. Nearest Neighbor algorithm was employed as the prediction engine. The peptides
around the phosphorylation sites with a fixed length of thirteen amino acid residues were extracted via a sliding window
along the protein chains concerned. Each of such peptides was coded into a vector with 6,072 features, derived from
Amino Acid Index (AAIndex) database, for the classification/detection. Incremental Feature Selection, a feature selection
algorithm based on the Maximum Relevancy Minimum Redundancy (mRMR) method was used to select a compact fea-
ture set for a further improvement of the classification performance. Three predictors were established for identifying the
three types of phosphorylation sites, achieving the overall accuracies of 66.64%, 66.11% and 66.69%, respectively. These
rates were obtained by rigorous jackknife cross-validation tests.
Keywords: Data mining, phosphorylation, aaindex; mRMR, machine learning approach.

INTRODUCTIONS also act as evidences that protein phosphorylation is a gen-

Phosphorylation is one of the most important post-
translational modifications which play a fundamental role in
most of the cellular regulatory pathways [1-3]. Phosphoryla-
tion of a protein is catalyzed by protein kinases (PKs) which
exhibit different preferences in the modification of substrate
residues. The phosphorylation of hydroxyl group of the side
chains of serine, threonine, and tyrosine plays extensive roles
in living cells for signal transduction cascades [4, 5]. It is in
demand to identify the substrates accompanied with their
phosphorylation sites in large-scale phosphoproteome, for it
may aid researchers on disease analysis and drug design [6].
The existing ways to identify the phosphorylation sites
include the experimental and computational methods. In the
past few years, much effort was made to analyze bacterial
phosphoproteomes with 2D gels followed by mass spectro-
metric identification of [32P]-labeled spots [7] and titanium
oxide chromatography, followed by ion trap-Fourier trans-
form ion cyclotron resonance mass spectrometer [8]. These
studies significantly increase the number of bacterial proteins
which can be phosphorylated on Ser/Thr/Tyr residues, and
*Address correspondence to this author at the CAS-MPG Partner Institute
for Computational Biology, Shanghai Institutes for Biological Sciences
(SIBS), Chinese Academy of Sciences (CAS), 320 Yueyang Road, Shanghai
200031, China; Tel: 86-21-54920498; Fax: 86-21-54920451;
E-mail: jfsamery@gmail.com
1875-5305/12 $58.00+.00
eral and fundamental regulatory process. In yeast, it was
reported that the use of proteome chip provided a first-
generation phosphorylation map [9]. High-throughput eu-
karyotes studies such as recent large-scale analysis of the
human phosphoproteome by quantitative mass spectrometry,
in which the time courses of more than 6,600 protein phos-
phorylation sites in HeLa cells were measured that response
to epidermal growth factor (EGF) stimulation, enables re-
searchers to study biological systems from a global perspec-
tive [10].
However, experimental detection of protein phosphoryla-
tion sites is time-consuming and often limited by the avail-
ability and optimization of enzymatic reactions. The predic-
tion of phosphorylation sites with their specific kinase using
computational approaches based on their primary sequences
could be of help in this regard, because computational meth-
ods can provide automatic and fast annotations, which can
hopefully be in turn used as a reference and guideline for
conducting experiments and for the interpretation of phos-
phoproteomic data. In view of this, many in silico algorithms
have been proposed. Meanwhile, the web servers enabling
users to submit and analyze their own data have also been
developed to facilitate biological experiments. For example,
the consensus sequences, motifs, function modules and spe-
cific residues have been adopted in the phosphorylation sites
prediction [11], other predictors including KinasePhos, based
on the profile hidden Markov model [12], PPSP adopting the
© 2012 Bentham Science Publishers
A Novel Predictor for Phosphorylation Site
Bayesian Discriminant Method [13], PredPhospho and
PHOSIDA based on Support Vector Machine (SVM) [14,
15], information-Entropy based Phosphorylation Prediction
for the automatic detection of potential phosphorylation sites
[16], and NetworKIN which is a database of predicted
kinase–substrate relationship based on the latest human
phosphoproteome and protein association network [17].
In the computational biology and bioinformatics area,
machine learning and data mining methods have been widely
used by many researchers who have made great efforts to
develop useful algorithms and software to investigate differ-
ent biological problems such as protein post-translation
modification, protein subcellular locations, protein-DNA
interaction and so on [18-29].
In this study, we developed a novel sequence-based
method for predicting phosphorylation site based on machine
learning approach (NNA, Nearest Neighbor Algorithm)
combining with feature selection (IFS and FFS based on
mRMR [30]) using an independent dataset. Three independ-
ent classifiers were developed for three types of phosphory-
lation sites: serine, threonine and tyrosine, respectively. Fea-
ture selection was used in our study, not only for improving
the predictor's performance, but also for analyzing the fac-
tors affecting the appearance of phosphorylation site. Finally,
we obtained overall success rates of 66.64%, 66.11% and
66.69% respectively for serine, threonine and tyrosine phos-
phorylation site predictions. Our analysis shows that some
particular positions (including the phosphorylation site) play
important roles during the phosphorylation process. It also
shows that hydrophobicity and the secondary structure of
amino acids in the flanking sequences are important for the
phosphorylation process.
MATERIALS AND METHODS
Benchmark Datasets
In this paper, the dataset was built from the database
UniProtKB/Swiss-Prot (Release 55.4). The method of how
to determine phosphorylated site in UniProtKB database was
described in UniProt knowledgebase user manual. According
to the manual, the phosphorylation information was listed in
the MOD_RES key feature lines of FT field. The informa-
tion contained phosphorylated residue name, and its site
number. For example, if a phosphorylation information re-
cord was “FT MOD_RES 198 Phosphoserine”, it means that
the 198th residue in a sequence was marked as a phosphory-
lated site by experiments and the residue name was serine.
However, when there was a “Potential” or “By similarity”
mark, it was not accepted as a phosphorylated site in this
study because these marks indicate that the result was not
proved by experiments but by some sort of deduction. In a
sequence, the following residues were possible to be the
phosphorylated site: Serine, Threonine, Tyrosine, Aspartate,
Histidine, or Cysteine, and, more rarely, Arginine. In this
paper, only serine (S), threonine (T) and tyrosine (Y) were
considered because the number of other phosphorylated resi-
dues was too few to have statistical significance.
At first, 366,226 sequences were extracted from the data-
base with their accession numbers. In case a protein had
more than one accession number, only the first accession
Protein & Peptide Letters, 2012, Vol. 19, No. 1 71
number was accepted. After removing redundancy, 71,637
proteins were obtained and these were further refined by
only selecting those that had the phosphorylated sites con-
cerned. Eventually, we obtained the following three datasets:
(1) set-S containing 4,652 proteins with phosphorylated site
S; (2) set-T containing 1,817 proteins with phosphorylated
site T; (3) set-Y containing 771 proteins with phosphorylated
site Y.
From the three protein sets, the corresponding peptide
fragment sets were generated according to the following pro-
cedures. If the protein set was set-S, all peptides that contain-
ing S would be picked out through a sliding window [31]
along each of the protein sequences concerned. Considering
the computational complexity, the peptide fragments were
composed of 13 residues with 6 residues upstream and 6
residues downstream of S located at the center. Similar op-
erations were taken for T and Y, respectively. Thus, we ob-
tained three peptide fragment datasets as denoted by pep-S,
pep-T and pep-Y, respectively. The confirmed phosphoryla-
tion peptides were assigned as positive samples; while the
rest were assigned as negative ones.
Since the size of the negative samples thus obtained was
much larger than that of the positive samples, the robustness
of the prediction model could be affected by the unbalanced
data size between the positive and negative samples. To
avoid this situation, we kept all the positive samples while
the negative samples were randomly selected from the origi-
nal data set until the number of the negative samples was
twice as that of the positive ones. Finally, we obtained (1)
pep-S containing 13,855 positive samples (Online Support-
ing Information 1A) and 27,710 negative samples (Online
Supporting Information 1B), (2) pep-T containing 2,812
positive samples (Online Supporting Information 2A) and
5624 negative samples (Online Supporting Information 2B),
and (3) pep-Y containing 1,259 positive samples (Online
Supporting Information 3A) and 2,518 negative samples
(Online Supporting Information 3B).
Feature Vector Construction
AAIndex [32, 33] contains hundreds of physiochemical
or biological amino acid properties and each index represents
a kind of amino acid properties, presented in a form of nu-
meric matrixes. These features covered different aspect of
amino acid characteristics, including alpha and turn propen-
sities, beta propensity, composition, hydrophobicity, phys-
icochemical properties and so on. AAindex consists of two
sections: AAindex1 for the collection of published amino
acid indices and AAindex2 for the collection of published
amino acid mutation matrices. Here AAIndex1 section was
used. So far, several previous studies have already used
AAIndex properties as parts of the features they considered
when applied machine learning approach to biological stud-
ies [34, 35]. AAIndex release 8.0, containing 562 indices,
was used to represent our peptide samples. 506 of them were
selected because indices containing missing values or am-
biguous annotations were excluded. Hence a 12-residue pep-
tide sample could be encoded by a 506 12 = 6072D (di-
mensional) vector, as formulated by
(
P = f0 , f1 , , fi , , fn ) ( i = 0, 1, , 6071 ) (1)