Protein Tyrosine Sulfation Sites
WEN-CHI CHANG,1,2* TZONG-YI LEE,2* DRAY-MING SHIEN,3,5JUSTIN BO-KAI HSU,2JORNG-TZONG HORNG,3,6 PO-CHIANG HSU,2TING-YUAN WANG,2HSIEN-DA HUANG,1,2RONG-LONG PAN4
1
Department of Biological Science and Technology, National Chiao Tung University, Hsin-Chu, Taiwan
2
Institute of Bioinformatics and Systems Biology, National Chiao Tung University, Hsin-Chu, Taiwan
3
Department of Computer Science and Information Engineering, National Central University, Chung-Li 320, Taiwan
4Institute of Bioinformatics and Structural Biology, College of Life Sciences,
National Tsing Hua University, Hsin-Chu, Taiwan 5
Department of Electronic Engineering, Chin Min Institute of Technology, Miao-Li, Taiwan 6
Department of Bioinformatics, Asia University, Taichung, Taiwan Received 31 October 2008; Revised 21 January 2009; Accepted 2 February 2009
DOI 10.1002/jcc.21258
Published online 16 April 2009 in Wiley InterScience (www.interscience.wiley.com).
Abstract: Tyrosine sulfation is a post-translational modification of many secreted and membrane-bound proteins. It governs protein-protein interactions that are involved in leukocyte adhesion, hemostasis, and chemokine signaling. However, the intrinsic feature of sulfated protein remains elusive and remains to be delineated. This investigation presents SulfoSite, which is a computational method based on a support vector machine (SVM) for predicting pro-tein sulfotyrosine sites. The approach was developed to consider structural information such as concerning the sec-ondary structure and solvent accessibility of amino acids that surround the sulfotyrosine sites. One hundred sixty-two experimentally verified tyrosine sulfation sites were identified using UniProtKB/SwissProt release 53.0. The results of a five-fold cross-validation evaluation suggest that the accessibility of the solvent around the sulfotyrosine sites contributes substantially to predictive accuracy. The SVM classifier can achieve an accuracy of 94.2% in five-fold cross validation when sequence positional weighted matrix (PWM) is coupled with values of the accessible sur-face area (ASA). The proposed method significantly outperforms previous methods for accurately predicting the location of tyrosine sulfation sites.
q 2009 Wiley Periodicals, Inc. J Comput Chem 30: 2526–2537, 2009
Key words: protein; sulfation; prediction
Introduction
Numerous post-translational modifications (PTMs) of proteins provide the proteome with structural and functional diversity, and regulate cellular plasticity and dynamics. Tyrosine sulfation is one of the most common PTMs in secreted and transmem-brane proteins, and has been experimentally demonstrated to be
essential to extracellular protein–protein interactions.1,2
Approxi-mately 1% of all tyrosine residues of the total proteins in an
or-ganism can be sulfated.3 Several proteins that are known to be
tyrosine-sulfated play important roles in physiological processes, and a direct link between protein function and tyrosine sulfation
has been established in many cases.4For example, sulfated
tyro-sines have an essential role in the immune response,5 such as
promote HIV infection of T-helper lymphocytes,1,6 leukocyte
rollin,7and a complementary cascade.8Sulfation is also involved
in the modulation of intracellular protein transportation, and the
regulation of the proteolytic process of proteins.9Tyrosine
sulfa-tion occurs when tyrosylprotein sulfotransferases (TPSTs)
cata-lyze the transfer of a negatively charged sulfate from 30
-phos-phoadenosine-50-phosphosulfate (PAPS) to the hydroxyl group
of tyrosine residue on a polypeptide.10Two tyrosylprotein
sulfo-Additional Supporting Information may be found in the online version of this article.
*Both the authors contributed equally to this work.
Correspondence to: H.-D. Huang; e-mail: bryan@mail.nctu.edu.tw or R.-L. Pan; e-mail: rlpan@life.nthu.edu.tw
Contract/grant sponsor: National Science Council of the Republic of China; contract/grant number: NSC 97-2811-B-009-001
Contract/grant sponsor: National Research Program for Genomic Medi-cine (NRPGM), Taiwan
transferases (TPST1 and TPST2) have been identified.10–13 The
use of synthetic peptide acceptorsin vitro suggests that the two
TPST proteins prefer different substrates.12 For instance,
pepti-des modeled on the N-terminus of P-selectin glycoprotein ligand-1 are sulfated by the two isoenzymes with equal
effi-ciency.14 Furthermore, peptides that are modeled on sulfation
sites of human C4-a-chain and heparin cofactor II are sulfated
more efficiently by TPST1 than by TPST2.12In contrast, Danan
et al. demonstrated that TPST2 is more catalytically efficient
than TPST1 in sulfating CCR8 substrates.15The two
tyrosylpro-tein sulfotransferases, TPST-1 and TPST-2, have overlapping
but not identical substrate specificitiesin vitro and in vivo.2,12,15
However, the relative abundances and distribution of the two isoenzymes have not yet been examined because of lack of
suit-able analytical reagents and tools.10,12
Tyrosine-sulfated protein is important to cellular control, but no clear-cut acceptor sequence of TPSTs, which can be used to predict sulfotyrosine sites. Because of labile sulfotyrosine, the characterization of this PTM has been impeded by the limitation of general, unambiguous methods for its site determination. Eliminating numerous false positive tyrosine sulfations, Rosen-quist and Nicholas discovered a test for acidic amino acids close to the target tyrosine, which yielded a good filtering
crite-rion.16,17However, no universal features have yet been identified
for use in the accurate prediction of sulfation sites. To increase the predictive accuracy and to reduce the number of false-posi-tives, Yu et al. incorporated context-based rules and logical fil-ters to develop a so-called position-specific scoring matrix
(PSSM) to predict tyrosine sulfation sites.18 Unfortunately, the
capacity to find new or unseen PTM patterns is absent due to
the static behavior of PSSMs.12,18 Additionally, the methods
demonstrated above without a user-friendly interface.
Sulfina-tor19was developed to predict sulfotyrosine sites based on four
Hidden Markov Models. Despite the high predictive accuracy of Sulfinator’s validation test set, it cannot be guaranteed to iden-tify new sulfated tyrosine sites from all of the protein sequence
databases.12 For instance, variable sulfotyrosine in extracellular
class II leucine-rich repeat (LRR) proteins were characterized by
mass spectrometry, and, Sulfinator19 does not accurately predict
tyrosine sulfation in this class of proteins.3Therefore, a
compu-tational tool to predict accurately tyrosine sulfation sites in protein sequences is crucial.
This work develops a new method for identifying protein sul-fotyrosine residues. The studies cited earlier have been only pro-posed to analyze the consensus sequences around sulfotyrosine sites and did not take consider other information. Most PTM sites are located at the surface of the protein, and the residues with larger Accessible Surface Areas (ASA) are regarded as surface
residues.20In this investigation, position-specific amino acid
fre-quencies and the structural property of the protein, in terms of the Accessible Surface Area (ASA), are employed to distinguish between sulfation and non-sulfation sites using a Support Vector
Machine (SVM).21 All tyrosine-sulfated proteins were collected
from UniProtKB/SwissProt (Release 53.0).22 Subsequently, the
experimental tyrosine-sulfated proteins were selected to reduce the quantity of the homologous sequences with a given window size and processed for SVM training. Suitable parameters were used to train SVM models. Five-fold cross-validation was
adopted to evaluate the accuracy of the models. The estimates made using the models reveal that the solvent accessibility around the sulfation sites can be used to improve their predictive accu-racy. Furthermore, during the independent test experiment, the proposed method achieves 84% predictive accuracy. The phos-phorylation and sulfation of tyrosine are well known to be iso-baric. Hence, differentiating modifications by phosphorylation from those by sulfation is generally difficult. However, the method herein effectively distinguishes sulfotyrosine from phos-photyrosine sites. SulfoSite is now available as a freely accessible web server at http://SulfoSite.mbc.nctu.edu.tw.
Materials and Methods
Various machine learning approaches (such as Linear Discrimi-nant Analysis (LDA), Support Vector Machine (SVM), and others) support the rapid collection, annotation, retrieval, compar-ison, and mining of specific features from large biological
data-sets, to solve structure–function problems.23–25A similar research
flow was also utilized in this study. Figure 1 presents the flow-chart of the proposed method. The method comprises three major steps: (i) collecting and preprocessing data, (ii) extraction of fea-tures, and (iii) creation and evaluation of models.
Collecting and Preprocessing Data
Collecting and Preprocessing Data to Construct and Test Models
UniProtKB/SwissProt22(release 53.0) maintains 269,293 protein
entries, of which 313 contain 732 residues, that are annotated as ‘‘sulfotyrosine’’ in the ‘‘MOD_RES’’ fields (and exhibit post-translational modification of a residue). However, only experi-mentally verified sulfotyrosine sites were collected. Potential sulfotyrosine sites for which the comments in UniProtKB/Swis-sProt’s contained the keywords of ‘‘by similarity,’’ ‘‘potential,’’ or ‘‘probable’’ were removed. As given in Table 1, 162 experi-mentally confirmed sulfotyrosine sites within 115 proteins that were extracted from UniProtKB/SwissProt release 53.0 were used for further preprocessing. Thereafter, the sulfotyrosine sites were extracted as positive sets, and the non-sulfotyrosine sites as negative sets from 115 tyrosine sulfated proteins.
Then, the 9-mer sequences (24 to 14), 11-mer sequences
(25 to 15), 13-mer sequences (26 to 16), 15-mer sequences
(27 to 17), and 17-mer sequences (28 to 18) of sulfotyrosine sites are extracted and constructed as training sets. The positive (1) dataset for training may contain some homologous sites of homologous proteins. When the test data were highly homolo-gous with the training data, the prediction accuracy was overes-timated. To avoid such overestimation, when two sulfotyrosine sites of the two proteins were in the corresponding positions in the alignment, only one was kept. Thus, non-homologous
posi-tive (1) data on high-quality tyrosine-sulfated sites were
obtained using different window sizes, as given in Table S1 (see Supporting Information). Nonhomologous positive and negative dataset were merged into a nonhomologous dataset. Moreover, the tyrosine-sulfated proteins in UniProtKB/SwissProt release 55.0, which were not available in release 53.0 were selected to
perform an independent test. The statistics of sulfation sites obtained from UniProtKB/SwissProt release 55.0 is given in Table S2 (see Supporting Information).
Collecting Data for Analysis of 3D Structure of Tyrosine-Sulfated Proteins
To elucidate the structural characteristics of sulfated protein, all matched 3D structures were downloaded from the protein dada
bank (PDB),26 as listed in Table S3 (see Supporting
Informa-tion). Table 2 summarizes structural information on the non-redundant subset of sulfated proteins in PDB. In the case of
ty-rosine, only nine hit proteins (CXAA_CONPE, CXAB_CONPE,
FIBG_HUMAN, GP1BA_HUMAN, HEP2_HUMAN, ITH1_
HIRME, ITH3_HIRME, ITHA_HIRME, and TRY1_HUMAN) contained sulfotyrosine, and only 13 sulfated sites had reliable structures, as shown in Table 2. The majority of these proteins are related to the thrombin inhibitor, the others are conotoxins and trypsin, as presented in Table S4 (see Supporting Informa-tion). All of them indicated that the absence of sulfation in the critical tyrosine influences biological activity (Table S4, see Supporting Information). Some proteins (small peptides) have
pharmacological applications; for example, conotoxins are
utilized in the selection of a lead compound in drug design. Figure 1. System flow of SulfoSite.
Extracting Features
After the nonhomologous training set of tyrosine sulfation sites had been constructed, three characteristics—sequences that round the sulfotyrosine sites, the secondary structure that sur-rounds the sulfotyrosine sites, and the solvent accessibility around the sulfotyrosine sites were extracted. However, the ex-perimental ASA and SS information about the sulfotyrosine sites
of known 3D structures in the protein data bank (PDB)26 is not
enough to be examined (Table 1). Consequently, RVP-Net20,27
and PSIPRED28 were applied to predict the Accessible Surface
Area (ASA) and secondary structure (SS), respectively, for each residue of a protein sequence. RVP-Net directly predicts ASA without making exposure categories and achieves a mean
abso-lute error better than 19% mean.27PSIPRED28is a highly
accu-rate method for predicting the secondary structure, combining two feed-forward neural networks to analyze the output obtained
using Position Specific Iterated-BLAST (PSI-BLAST).29
PSIPRED 2.0 achieved an average accuracy of 80.6% across all 40 submitted target domains with no obvious sequence similarity to structures that are present in PDB, making PSIPRED better
than all other evaluated methods.30 The output of PSIPRED
includes three symbols ‘‘H,’’ ‘‘C,’’ and ‘‘E’’ which stand for
helix, coil, andb-sheet respectively. Before SVM models are
cre-ated, the extracted features should be encoded as real values except the ASA values. The two sequences around the
sulfotyro-sine site (position 0) and the 9-mer sequences (24 to 14), 11-mer
sequences (25 to 15), 13-mer sequences (26 to 16), 15-mer sequences (27 to 17), and 17-mer sequences (28 to 18) of the tyrosine sulfation sites were encoded in three ways—two for the sequence and one for the secondary structure. As presented in Table S5 (see Supporting Information), an orthogonal binary coding scheme was adopted to transform amino acids into nu-meric vectors, as in so-called 20-dimensional vector coding. For example, alanine was encoded as ‘‘10000000000000000000’’; Table 1. The Statistics of Sulfation Sites Obtained from
UniProtKB/SwissProt Release 53.0. Number of proteins Number of sulfotyrosine sites Tyrosine sulfated protein 313 732a
Experimental tyrosine sulfated protein 115 162b
Experimental secreted tyrosine sulfated protein
103 138b
Experimental membrane-bound tyrosine sulfated protein
8 17b Experimental Tyrosine sulfated protein
annotated in PDB
9 12
Nonhomologous experimental sulfation sites in SulfoSite
– 69c
a
Those sites annotated as ‘‘by similarity’’, ‘‘potential,’’ or ‘‘probable’’ are included.
b
Those sites annotated as ‘‘by similarity’’, ‘‘potential,’’ or ‘‘probable’’ are not included.
c
On the basis of window size 4.
Table 2 . Hits between sulfate d proteins fro m U niProtK B/SwissP rot 53.0 an d PDB. Swi ssPro t ID PDB ID Ch ain PDB cover age in SwissPr ot The position of sul fated tyros ine in Swi ssProt protein SS of sul fated tyrosin e in crystal struct ure ASA percentage of sulfate d tyrosin e (posi tion 5 0) in cry stal struct ure 2 6 2 5 2 4 2 3 2 2 2 10 1 1 1 2 1 3 1 4 1 5 1 6 CXAA _CON PE 1PEN A 1–16 Y15 coil 0.26 0.74 0.80 0.39 0.73 0.88 0.86 0.50 – – – – – CXAB _CO NPE 1AKG A 45–6 0 Y59 coil 0.28 0.90 0.73 0.41 0.62 0.65 0.83 0.54 – – – – – FIB G_HUMA N 2 H W L P 439– 452 Y444 coil – 0.11 0.24 0.39 0.37 0.66 0.51 0.11 0.45 0.29 0.90 0.97 0.78 Y448 coil 0.37 0.66 0.51 0.11 0.45 0.29 0.90 0.97 0.78 0.87 0.89 – – G P1BA_H UMAN 1GWB B 16–2 96 Y294 coil 0.52 0.20 0.00 0.15 0.05 0.15 0.07 0.35 0.51 – – – – 1P8V A 17–2 95 Y292 coil 0.97 0.50 0.31 0.65 0.58 0.30 0.17 0.12 0.62 0.75 – – – Y295 coil 0.65 0.58 0.30 0.17 0.12 0.62 0.75 – – – – – – H EP2_HUM A N 1JMO A 20–4 99 Y79 coil 0.33 0.54 0.33 0.50 0.33 0.31 0.26 0.02 0.45 0.20 0.88 0.61 0.12 Y92 helix 0.18 0.32 0.72 0.19 0.91 0.47 0.95 0.78 0.67 0.52 0.92 0.79 0.89 ITH 1_HIR ME 1B CU I 1–65 Y63 helix 0.40 0.78 0.08 0.24 0.89 0.67 0.28 0.64 – – – – – ITH 3_HIR ME 1W7G I 60–7 1 Y70 coil 0.38 0.68 0.18 0.42 – 0.89 0.38 – – – – – – ITH A_HIRM E 1A46 I 54–6 4 Y63 coil 0.46 0.25 0.11 0.38 0.29 0.37 0.41 0.60 – – – – – TR Y1_HUM AN 1TRN A 24–2 47 Y154 coil 0.03 0.25 0.54 0.67 0.77 0.45 0.53 0.13 0.47 0.37 0.03 0.11 0.01 More details are shown in Table S3 (se e Support ing Inform ation). If SwissP rot_ID hits mo re than one PDB _ID, onl y one is shown in this tabl e.
cysteine encoded as ‘‘01000000000000000000,’’ and so on. The number of feature vectors that correspond to the flanking amino
acids that surround the sulfated site was (2n1 1) 3 20. Values
ofn from four to eight were used to determine the optimal
win-dow length. The 20-dimensional vector for protein sequence encoding (with each amino acid mapped to a 20-dimensional vector) was named SEQ(20-d). Additionally, the positional weighted matrix (PWM) of amino acids around the sulfated sites was determined for tyrosine using nonhomologous training data. The positional weighted matrix (PWM) specified the relative fre-quency of amino acids in the sulfated sites, and was used to encode the fragment sequences, namely SEQ(PWM) (Table S5 in Supporting Information). Whenever feature values were miss-ing (such as in case of tyrosine residues near the end of the pep-tide), they were discarded. Furthermore, the full-length protein sequences with sulfated sites were inputted to PSIPRED to ver-ify the secondary structure of all residues. The orthogonal binary coding scheme was used to transform the three terms that spec-ify the secondary structure into numeric vectors. For instance, helix was encoded as ‘‘001,’’ sheet was encoded as ‘‘010,’’ and coil was encoded as ‘‘100.’’ The 3D vector that encoded the sec-ondary structure (SS) of the protein (each secsec-ondary structure was mapped to a 3D vector) was named SS. Figure S1 (see Supporting Information) presents a diagram of the feature dimensions.
Creation and Evaluation of Model
SVM is a machine learning method, which has been utilized to solve pattern identification problems with clear connections to
the underlying statistical learning theory.21 The principle of
SVM is to map input vectors to a higher dimensional space
where a maximal separating hyperplane is defined. Two parallel hyperplanes are constructed on each side of the hyperplane that separates the data into two groups. The separating hyperplane is that which maximizes the distance between the two parallel hyperplanes. Furthermore, SVM can solve the classification
problem when the number of training data is very small.31LDA
is a simpler classifier than SVM that is adopted to solve
biologi-cal structure–function problems.32–34People may argue with our
reasons for selecting SVM as the method herein. Therefore, the classification performance of LDA and SVM were compared ini-tially. As given in Table S6 (see Supporting Information), the models trained by SVM greatly outperform those trained using LDA based on the same training and validation data sets. Accordingly, this work employed protein sequences, profiles of Accessible Surface Area (ASA) and secondary structure (SS) in the support vector machine (SVM) for training the models of
sulfotyrosine site prediction. An SVM library called LIBSVM,35
was applied to train the predictive models and a radial basis function (RBF) was selected as the SVM kernel function in this study. The experimental sulfotyrosine sites were defined as the positive dataset, while all other nonsulfated tyrosine sites in the sulfated proteins were treated as the negative dataset. K-fold cross-validation was exploited to evaluate the predictive per-formance of the models. The sizes of the positive and negative datasets were made equal during the cross-validation processes. After the models were trained, whether they were the models are appropriate had to be determined. The performance of SVM was measured in a five-fold cross-validation analysis, in which each dataset was divided into five parts: four parts were used for model learning (training); the remaining one was used for vali-dation (testing). Four performance measures were employed; they were precision (Pr), sensitivity (Sn), specificity (Sp), and Table 3. The Statistics of Sequence Logo, Average Accessible Surface Srea (ASA), and Secondary Structure
(SS) Surrounding the Sulfotyrosine Sites (Position5 0) in UniProtKB/SwissProt Release 53.0.
Sequence logos Accessible surface area (ASA) Secondary structure Entire experimental tyrosine sulfated sites Nonhomologous tyrosine sulfated sites Nonsulfated tyrosine sites
accuracy (Acc), defined later. The Matthews correlation coeffi-cient (MCC) was used to measure the classification performance in the SVM training process to accommodate unbalanced data-sets. TP, TN, FP, and FN are true positive, true negative, false positive, and false negative predictions, respectively.
PrecisionðPrÞ ¼ TP ðTP þ FPÞ; Sensitivity ðSnÞ ¼ TP ðTP þ FNÞ; SpecificityðSpÞ ¼ TN ðTN þ FPÞ; AccuracyðAccÞ ¼ ðTP þ TNÞ ðTP þ FP þ TN þ FNÞ; MCC¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiðTP3TNÞ ðFN3FPÞ ðTP þ FNÞ3ðTN þ FPÞ3ðTP þ FPÞ3ðTN þ FNÞ p
Several parameters of the models, the length of the sequence that surround the sulfotyrosine site, the SVM cost and the gamma values were optimized to maximize predictive accuracy. Additionally, receiver operating characteristic (ROC) curves for
different window sizes and features were plotted using
LIBSVM.35 Finally, the parameters of the trained model with
the highest predictive accuracy from each dataset were adopted to provide the prediction service on the web. SulfoSite was implemented in PHP, PERL, and hosted by an Apache web server running on a RedHat Linux system.
Results and Discussion
Characteristics of Sulfated Protein
First, all sulfated proteins from UniProtKB/SwissProt22 (release
53.0) were collected; 115 of them were experimentally anno-tated. As indicated in Table 1, the numbers of secreted and membrane-bound tyrosine sulfated proteins were 103 and eight, which contained 138 and 17 sulfotyrosine sites, respectively.
Table 3 summarizes the sequence logos,36 average ASA, and
secondary structure formed from the 9-mer sulfated and the 9-mer non-sulfated sites in the constructed data set. As the
flank-ing sequences (position 24 to 14) of the sulfotyrosine sites
(position 0) are graphically visualized as sequence logos36of the
protein sequence and the secondary structure (SS), examining the sequence logos reveals that the sequences that surround the sulfated sites are more acidic and more strongly conserved than those around non-sulfated sites. Furthermore, Accessible Surface Area (ASA) is higher at the sulfotyrosine site than at the flank sites.
On the basis of the 3D structures from the protein data bank
(PDB),26Table 2 demonstrates that two sulfotyrosine residues in
PDB proteins were observed in helical regions, while the others were observed in the loop. Furthermore, the ASA percentages of some sulfotyrosine residues exceeded those around the sulfated site, indicating that the sulfotyrosine is usually on the protein surface. Additionally, the locations of sulfotyrosines in PDB teins classified them into three broad types (main chain of pro-tein complex, small peptide, or inhibitor), as displayed in Figure S2 (see Supporting Information). The figure also indicates the sulfotyrosine residues located on the structure surface. The 3D structure is consistent with the sequence analysis that is given in Table 3, potentially explaining why the ASA feature is critical to predicting the sites of sulfotyrosine.
Algorithm
Predictive Performance of Cross Validation–Individual Feature
Although SVM is effectively ignores irrelevant features in high dimensional feature sets, proper attribute selection may further increase the accuracy of the SVM algorithm. In previous studies, methods for predicting sulfotyrosine sites have focused on the conservation of the protein sequence: an optimal discriminative set of conserved protein sequences and structural properties had
not been identified.19The predictive accuracy was used to
evalu-ate four individual properties (SEQ(20-d), SEQ(PWM), ASA, Figure 2. Performance of the models trained with various features
and different window sizes in five-fold cross-validation. (A) The predictive accuracies of the models trained with SEQ(20-d), SEQ(PWM), ASA, and SS based on various window sizes. (B) The predictive accuracies of the models trained with SEQ(20-d), SEQ(PWM), ASA, SS, SEQ(20-d)1ASA, SEQ(20-d)1SS, SEQ (20-d)1ASA1SS, SEQ(PWM)1ASA, SEQ(PWM)1SS, and SEQ (PWM)1ASA1SS based on various window sizes. Abbreviation: SEQ(20-d), protein sequence encode by 20-dimentional method; ASA, Accessible Surface Area; SS, secondary structure; SEQ(PWM), protein sequence encode by positional weighted matrix (PWM) method. [Color figure can be viewed in the online issue, which is available at www.interscience.wiley.com.]
and SS) of the residues to determine the optimal subset of such properties. The average predictive accuracy of individual attrib-utes, except SEQ(PWM), was low. As shown in Figure 2A and Table 4, that of SEQ(PWM) was 89.58%, that of SEQ(20-d) was 84.93%, that of ASA was 76.24%, and that of SS was only 72.35%. A model trained with a window size 9 (24 to 14) based on SEQ(PWM) performed best in predicting sulfotyrosine (91.30%).
Predictive Performance of Cross Validation–Combinations of Feature Sets
The predictive performance of models trained with the combined attributes was also estimated. The average prediction accuracy increased to 85.62% when the SEQ(20-d) was coupled with ASA, as given in Table 4. However, the predictive accuracy of SEQ(20-d)1 ASA was not satisfactory. Since the number of dimensions in SEQ(20-d) was 20 times that in ASA, the weight of SEQ(20-d) exceeded that of ASA in the prediction of the sites of sulfated tyrosine. Thus, the predictive performance was dominated by the sequence feature. The weight of ASA was
increased five-fold, 10–fold, and 20-fold, and the average accu-racy did not significantly change for SEQ(20-d)1ASA_5X (five-fold), but improved to 86.06%, and 87.35% for SEQ(20-d)1ASA_10X (10-fold), and SEQ(20-d)1ASA_20X (20-fold), suggesting that ASA notably improved the predictive accuracy to 92.21% when its weight increased to 20-fold in window size
9 (24 to 14) (Table 4 and Fig. S3, see Supporting Information).
For this reason, the weight of ASA is important in predicting the sites of sulfotyrosine. However, the number of dimensions in SEQ(PWM) equaled that of to ASA, balancing the weights to enable sulfated and non-sulfated tyrosine to be classified.
Although the average accuracy of SEQ(PWM)1ASA was not
better than that of SEQ(PWM), a model trained with
SEQ(PWM)1ASA with window size 9 (24 to 14) was the
most accurate (94.20%), as given in Figure 2B and Table 4. The AUC (area under curve) of the ROC curve was 0.9702 in the model that had been trained with SEQ(PWM)1ASA in window size 9 (24 to 14), which is the best value among all of the models, as given in Figure S4 (see Supporting Information). Additionally, the secondary structure (SS) feature did not mark-edly improve the predictive accuracy. The best predictive model
Table 5. The Parameters and Predictive Performance of the Trained Models with Different Features Which Achieve the Highest Accuracy.
Sulfated residue
Number of positive
training set Training features Window size Parameters Pr Sn Sp Acc MCC Tyrosine 69 SEQ(PWM)1ASA 24 to 14 Ca5 32768, Gb5 0.00048828125 95.52 92.75 95.65 94.20 0.88
aCost value. bGamma value
Table 4. Prediction Accuracy of the Models Trained with Various Parameters and Different Window Size in Five-Fold Cross-Validation. Training features Window size Average accuracy 9 11 13 15 17 SEQ(20-d) 87.13 86.03 86.36 82.45 82.69 84.93 SEQ(PWM) 91.30 88.97 90.91 87.27 89.42 89.58 ASA 71.74 77.94 75.45 78.18 77.88 76.24 SS 69.57 72.79 64.55 82.73 72.12 72.35 SEQ(20-d)1ASA 90.10 82.35 85.45 83.64 86.54 85.62 SEQ(PWM)1ASA 94.20b 85.29 88.18 84.55 86.54 87.75 SEQ(20-d)1SS 87.68 85.29 83.64 83.64 77.88 83.63 SEQ(PWM)1SS 88.41 89.71 84.55 79.09 85.58 85.47 SEQ(20-d)1ASA1SS 84.78 85.29 87.27 84.55 82.69 84.92 SEQ(PWM)1ASA1SS 83.33 85.29 77.27 74.55 81.73 80.44 SEQ(20-d)1ASA_5Xa 87.66 87.82 77.69 86.15 88.52 85.57 SEQ(20-d)1ASA_10Xa 88.31 87.82 88.46 85.38 80.33 86.06 SEQ(20-d)1ASA_20Xa 92.21 85.9 82.31 86.15 90.16 87.35 Average accuracy 85.88 84.65 82.47 82.95 83.24
The ROC curves are given in Figure S4 (see Supporting Inforamation).
a
ASA_NX: The weight of ASA increase toN folds.
b
was trained using a window size 9 (24 to 14) in sulfotyrosine prediction. Table 5 presents the number of training data, training features, window size, SVM cost and gamma values, precision, sensitivity, specificity, accuracy, and MCC of the best trained model.
Predictive Performance of Balanced and Unbalanced Dataset
In cross-validation, balanced positive and negative datasets were utilized to train models. However, for this purpose, sampling a negative dataset absolutely randomly is difficult because the number of the negative sets is much more than those positive sets. To avoid the tendentiousness in negative set extraction and reduce the false positive rate, the negative set for five-fold cross validation was collected randomly. Its size was separately one, two, three, four, five, six, seven, eight, nine, and 10 times that of the positive set. These unbalanced datasets are for training only, and then testing is conducted with a dataset whose posi-tive/negative ratio is an estimate of the ratio in an unmodified
dataset (has a positive/negative ratio of about 1/40). The results suggest that higher accuracy and specificity were related to more negative sets, as given in Figure 3 and Table 6. However, the sensitivity was approximately inversely proportional to the num-ber of negative dataset (see Fig. 3). The accuracy and specificity
almost reached their respective maximum values (Ac5 96.55%
and Sp5 97.78%), when the ratio of the number of the positive
set to that of the negative set was to 1:8. At this ration, the sen-sitivity had dramatically decreased to 75.56% (see Fig. 3). A trade-off had to be made between specificity and sensitivity. Although the accuracy at the ratio 1:2 was lower than that at 1:8, the sensitively remained at 86.67% and also high accuracy and specificity were maintained. Furthermore, the MCC value was the highest. Therefore, to gain the higher prediction sensi-tivity and the lower false positive rate, the model that had been
trained with 1:2 unbalanced dataset based on SEQ(PWM)1ASA
(window size5 9) was selected as the best model herein.
Figure 3. The performance of balanced and unbalanced dataset based on SEQ(PWM)1ASA in window size 9 (24 to 14).
Table 6. The Accuracies of Balanced and Unbalanced Dataset. Training feature: SEQ(PWM)1ASA
Window size: 9 (24 to 14)
Positive dataset:Negative dataset ACC Se Sp MCC 1:1 91.03 86.67 91.12 0.72 1:2 95.86 86.67 96.32 0.87 1:3 94.48 86.67 94.85 0.82 1:4 94.48 75.56 95.59 0.77 1:5 94.48 75.56 95.59 0.77 1:6 95.17 75.56 97.06 0.74 1:7 95.17 75.56 97.06 0.74 1:8 96.55 75.56 97.78 0.86 1:9 96.55 75.56 97.78 0.86 1:10 95.87 64.44 97.78 0.78
Table 7. Independent test in Sulfinator1and SulfoSite.
SwissProt_ID Real Sulfotyrosine sites Sulfotyrosine predicted in Sulfinator Sulfotyrosine predicted in SulfoSite
CCKN_CANFA Y52 Y52 Y52
FMOD_BOVIN Y20, Y38, Y53, Y55, Y63, Y65 Y38, Y42, Y45, Y47, Y62, Y64 Y38, Y55 FMOD_HUMAN Y20, Y53, Y55 Y38, Y39, Y42, Y45, Y47 Y20, Y39 HIS1_HUMAN Y46, Y49, Y53, Y55 Y46, Y49 Y46, Y49
LUM_MOUSE Y20, Y21, Y23, Y30 – Y23
OMD_HUMAN Y39, Y416, Y417 – Y31, Y39, Y51
PSK1_ORYSI Y80, Y82 Y80 Y80, Y82
PSK1_ORYSJ Y80, Y82 Y80 Y80, Y82
PSK2_ORYSI Y110, Y112 – Y110, Y112
PSK2_ORYSJ Y110, Y112 – Y110, Y112
SIAL_HUMAN Y313, Y314 Y259, Y263, Y265, Y271, Y275, Y278, Y290, Y293, Y297, Y299
Y265, Y271, Y305
Pr – 23% 73%
Sn – 19% 52%
Sp – 80% 94%
Acc – 65% 84%
Testing
Predictive Performance of Independent Test by the Model Trained by SEQ(PWM)1ASA
Eleven sulfated proteins (CCKN_CANFA, FMOD_BOVIN,
FMOD_HUMAN, HIS1_HUMAN, LUM_MOUSE, OMD_
HUMAN, PSK1_ORYSI, PSK1_ORYSJ, PSK2_ORYSI,
PSK2_ORYSJ, SIAL_HUMAN) in UniProtKB/SwissProt release 55.0, which were not available in UniProtKB/SwissProt release 53.0 were used for an independent test in SulfoSite, as given in Table 7. The independent test set involved 31 experimentally sulfated tyrosines and 98 non-sulfated tyrosines. Table 7 reveals that the system herein can predict the number of 16 sulfotyro-sines (true positive), and only six predictions are false positive. The sensitivity and specificity of prediction are 52 and 94%, respectively. Generally, the performance in the independent test is just a little lower than those obtained in cross-validation, because that the over-fitting can usually not be completely
pre-vented. Although the predictive accuracy in the independent test is not as high as that in cross-validation, the performance dem-onstrated in Tables 7 and 8 is acceptable. The precision and ac-curacy in the independent test are 73 and 84%, respectively,
which are much better than test in Sulfinator19 (precision 5
23%, accuracy 5 65%), suggesting that combination of
sequence (encoded in PWM) and ASA is effective in sulfated protein prediction.
Predictive Performance of Model Trained Only With SEQ(PWM) in Independent Test
To confirm further the effectiveness of ASA in the prediction of sulfotyrosine sites, another model based on SEQ(PWM) with window size 9 (24 to 14) was trained using un-balanced
(posi-tive: negative 5 1:2) datasets. Significantly, our final model
(trained by SEQ(PWM)1ASA) is also better in the forecasting of sulfotyrosine sites in an independent test than is the Table 8. Comparison Between Sulfinator1and SulfoSite.
Sulfinator SulfoSite
Number of sulfated protein for training 125 115a Separating experimental and potential
sulfotyrosine sites in training dataset
No Yes
Nonhomologous dataset No Yesb
Features used for prediction Protein sequence Protein sequence, Accessible Surface Area Method Hidden Markov Model Support Vector Machine Cross-validation performance Acc5 98% Acc5 94.2 % Independent testc Acc5 65%
(Sn5 19%, Sp 5 80%, Pr 5 23%)
Acc5 84%
(Sn5 52%, Sp 5 94%, Pr 5 73%)
aExperimental tyrosine sulfated protein. Those sulfotyrosine annotated as ‘‘by similarity’’, ‘‘potential,’’ or ‘‘probable’’
are not included.
bThe number of nonhomologous dataset is shown in Table S1 in Supporting Information. c31 positive sites and 98 negative sites were tested in Sulfiator and SulfoSite.
Table 9. Comparison Between Final Model (SEQ(PWM)1ASA) and SEQ(PWM) Model with Independent Test.
SWISS-PROT_ID Real Sulfotyrosine sites The model trained by SEQ(PWM) The model trained by SEQ(PWM)1ASA
CCKN_CANFA Y52 Y52 Y52
FMOD_BOVIN Y20, Y38, Y53, Y55, Y63, Y65 Y42 Y38, Y55
FMOD_HUMAN Y20, Y53, Y55 Y39 Y20, Y39
HIS1_HUMAN Y46, Y49, Y53, Y55 Y49 Y46, Y49
LUM_MOUSE Y20, Y21, Y23, Y30 Y23 Y23
OMD_HUMAN Y39, Y416, Y417 Y31, Y39, Y51 Y31, Y39, Y51
PSK1_ORYSI Y80, Y82 Y80, Y82 Y80, Y82
PSK1_ORYSJ Y80, Y82 Y80, Y82 Y80, Y82
PSK2_ORYSI Y110, Y112 Y110, Y112 Y110, Y112 PSK2_ORYSJ Y110, Y112 Y110, Y112 Y110, Y112 SIAL_HUMAN Y313, Y314 Y265, Y271, Y305 Y265, Y271, Y305
Pr – 63% 73%
Sn – 39% 52%
Sp – 93% 94%
Acc – 80% 84%
SEQ(PWM) model, as shown in Table 9, revealing that the accuracy can be improved by combining protein sequence (encoded using the PWM method) and ASA, as given in Figure 2 and Tables 4 and 9.
Several Cases Studies to Evaluate Performance of SulfoSite
Recently, tyrosine sulfation has been found to be prevalent in
human chemokine receptors and important in human disease.37
Yu et al., demonstrates Tyr10 and Tyr18 of C-C chemokine
re-ceptor type 1 (CCR1) can be sulfated by TPST.2The CCR1
pro-tein sequence was inputted into our system to identify sulfotyro-sine sites. Subsequently, sulfated Tyr10 and Tyr18 of CCR1
were specifically detected, as shown in Figure S5 (see Support-ing Information). Furthermore, the PTM of proteins by phospho-rylation is well-known to be the most abundant type of cellular
regulation.38 The tyrosine phosphorylation of proteins in the
cytoplasm has critical roles in multiple biological processes.9,39
It is because phosphotyrosine and sulfotyrosine are isobaric, dis-tinguishing between phosphorylation and sulfation modifications on tyrosine residue is typically difficult, even though the sequence logos and ASA surrounding the sulfotyrosine and
phosphotyrosine sites are different.40 Table S7 (see Supporting
Information) shows the amino acid sequences in the flanking of sulfotyrosine are more consensus than phosphotyrosine. It also demonstrates that sulfotyrosine has a higher ASA percentage. Figure 4. Web interface of SulfoSite. [Color figure can be viewed in the online issue, which is
Moreover, 1973 experimental phosphotyrosine sites from Uni-ProtKB/SwissProt release 55.0 were used to evaluate our system. Largely phosphotyrosine sites cannot be miscalculated as sulfo-tyrosine sites in SulfoSite. The false positive rate is only 4.7%.
Implementation
As mentioned earlier, the best model was trained with sequence (PWM) and Accessible Surface Area in window size 9 (24 to 14) was selected to utilize in the sulfotyrosine prediction sever. To avoid the biased extraction of the negative set, the
un-bal-anced (positive: negative 5 1:2) trained model was applied in
SulfoSite web site. The SulfoSite web server provides user-friendly input and output interfaces, as shown in Figure 4. Sev-eral proteins in FASTA format could be inputted to the system. The predictions are presented in a diagramed that includes sul-fated position, flanking amino acids, and the ASA probability, which were predicted by RVP-Net (see Fig. 4). The training dataset and independent test used in SulfoSite also can be down-loaded for analysis.
Conclusion
Because of the levity of sulfotyrosine,2 the characterization of
protein sulfation modification has been hampered by the lack of specific, definite methods for its site determination. This investigation developed the SulfoSite web server, a high-per-formance sulfotyrosine predictor based on the SVMs method with protein sequence and Accessible Surface Area (ASA). The proposed system is comparable to several previous meth-ods,12,16,18,19 none of which consider structural information about the protein. Not only was the conservation of the protein sequence considered, but also the solvent accessibility and sec-ondary structure around the sulfotyrosine sites were tested. The 3D structure data (information on ASA and secondary struc-ture) for sulfation proteins were limited, so two effective tools,
RVP-Net20 and PSIPRED,28 were used to compute the ASA
percentage and secondary structure, respectively, based on
pro-tein sequence. Finally, the model trained by SEQ(PWM)1ASA
was selected as the best predictive model in the SulfoSite web site. The method herein is 94.2% accurate in sulfotyrosine pre-diction in five-fold cross-validation. Tables 7 and 8 compare
the performance of our approach with that of Sufinator.19 The
comparison demonstrates that SulfoSite has much better predic-tive performance than Sulfinator in an independent test. The use of more training data in SulfoSite may be responsible for its higher accuracy than Sulfinator. The nonhomologous dataset in SulfoSite may also mitigate the over-fitting problem in Sulfi-nator.
Several cases studies suggest that our method performs well not only in detecting sulfotyrosine sites but also in distinguish-ing sulfotyrosine from phosphotyrosine sites (Fig. S5 and Table S7 in Supporting Information). On the basis of the above, this developed system can facilitate the characterization of protein tyrosine sulfation.
Availability
The SulfoSite program is available at http://SulfoSite.mbc. nctu.edu.tw/.
References
1. Kehoe, J. W.; Bertozzi, C. R. Chem Biol 2000, 7, R57.
2. Yu, Y.; Hoffhines, A. J.; Moore, K. L.; Leary, J. A. Nat Methods 2007, 4, 583.
3. Onnerfjord, P.; Heathfield, T. F.; Heinegard, D. J Biol Chem 2004, 279, 26.
4. Seibert, C.; Sakmar, T. P. Biopolymers 2008, 90, 459.
5. Lin, H. C.; Tsai, K.; Chang, B. L.; Liu, J.; Young, M.; Hsu, W.; Louie, S.; Nicholas, H. B., Jr.; Rosenquist, G. L. Biochem Biophys Res Commun 2003, 312, 1154.
6. Choe, H.; Li, W.; Wright, P. L.; Vasilieva, N.; Venturi, M.; Huang, C. C.; Grundner, C.; Dorfman, T.; Zwick, M. B.; Wang, L.; Rosen-berg, E. S.; Kwong, P. D.; Burton, D. R.; Robinson, J. E.; Sodroski, J. G.; Farzan, M. Cell 2003, 114, 161.
7. Bernimoulin, M. P.; Zeng, X. L.; Abbal, C.; Giraud, S.; Martinez, M.; Michielin, O.; Schapira, M.; Spertini, O. J Biol Chem 2003, 278, 37. 8. Gao, J.; Choe, H.; Bota, D.; Wright, P. L.; Gerard, C.; Gerard, N. P.
J Biol Chem 2003, 278, 37902.
9. Zhang, Y.; Jiang, H.; Go, E. P.; Desaire, H. J Am Soc Mass Spec-trom 2006, 17, 1282.
10. Moore, K. L. J Biol Chem 2003, 278, 24243.
11. Beisswanger, R.; Corbeil, D.; Vannier, C.; Thiele, C.; Dohrmann, U.; Kellner, R.; Ashman, K.; Niehrs, C.; Huttner, W. B. Proc Natl Acad Sci USA 1998, 95, 11134.
12. Monigatti, F.; Hekking, B.; Steen, H. Biochim Biophys Acta 2006, 1764, 1904.
13. Ouyang, Y. B.; Moore, K. L. J Biol Chem 1998, 273, 24770. 14. Wilkins, P. P.; Moore, K. L.; McEver, R. P.; Cummings, R. D.
J Biol Chem 1995, 270, 22677.
15. Danan, L. M.; Yu, Z.; Hoffhines, A. J.; Moore, K. L.; Leary, J. A. J Am Soc Mass Spectrom 2008, 19, 1459.
16. Rosenquist, G. L.; Nicholas, H. B., Jr. Protein Sci 1993, 2, 215. 17. Bundgaard, J. R.; Vuust, J.; Rehfeld, J. F. J Biol Chem 1997, 272,
21700.
18. Yu, K. M.; Liu, J.; Moy, R.; Lin, H. C.; Nicholas, H. B., Jr.; Rosen-quist, G. L. Endocrine 2002, 19, 333.
19. Monigatti, F.; Gasteiger, E.; Bairoch, A.; Jung, E. Bioinformatics 2002, 18, 769.
20. Ahmad, S.; Gromiha, M. M.; Sarai, A. Bioinformatics 2003, 19, 1849. 21. Vapnik, V. N. The Nature of Statistical Learning Theory; New York:
Springer-Verlag, 1995.
22. Boeckmann, B.; Bairoch, A.; Apweiler, R.; Blatter, M. C.; Estreicher, A.; Gasteiger, E.; Martin, M. J.; Michoud, K.; O’Donovan, C.; Phan, I.; Pilbout, S.; Schneider, M. Nucleic Acids Res 2003, 31, 365.
23. Gonzalez-Diaz, H.; Gonzalez-Diaz, Y.; Santana, L.; Ubeira, F. M.; Uriarte, E. Proteomics 2008, 8, 750.
24. Gonzalez-Diaz, H.; Vilar, S.; Santana, L.; Uriarte, E. Curr Topics Med Chem 2007, 7, 15.
25. Gonzalez-Diaz, H.; Molina, R.; Uriarte, E. FEBS Lett 2005, 579, 4297.
26. Deshpande, N.; Addess, K. J.; Bluhm, W. F.; Merino-Ott, J. C.; Townsend-Merino, W.; Zhang, Q.; Knezevich, C.; Xie, L.; Chen, L.; Feng, Z.; Green, R. K.; Flippen-Anderson, J. L.; Westbrook, J.; Berman, H. M.; Bourne, P. E. Nucleic Acids Res 2005, 33, D233; Database issue.
27. Ahmad, S.; Gromiha, M. M.; Sarai, A. Proteins 2003, 50, 629. 28. McGuffin, L. J.; Bryson, K.; Jones, D. T. Bioinformatics 2000, 16,
404.
29. Altschul, S. F.; Madden, T. L.; Schaffer, A. A.; Zhang, J.; Zhang, Z.; Miller, W.; Lipman, D. J. Nucleic Acids Res 1997, 25, 3389.
30. Bryson, K.; McGuffin, L. J.; Marsden, R. L.; Ward, J. J.; Sodhi, J. S.; Jones, D. T. Nucleic Acids Res 2005, 33, W36; Web Server issue.
31. Burges, C. J. C. Data Mining and Knowledge Discovery 1998, 2, 121.
32. Gonzalez-Diaz, H.; Aguero-Chapin, G.; Varona, J.; Molina, R.; Delogu, G.; Santana, L.; Uriarte, E.; Podda, G. J Comput Chem 2007, 28, 1049.
33. Gonzalez-Diaz, H.; Perez-Castillo, Y.; Podda, G.; Uriarte, E. J Com-put Chem 2007, 28, 1990.
34. Vilar, S.; Gonzalez-Diaz, H.; Santana, L.; Uriarte, E. J Comput Chem 2008, 29, 2613.
35. Chang, C. C.; Lin, C. J. In http://www.csie.ntu.edu.tw/~cjlin/libsvm 2001.
36. Schneider, T. D.; Stephens, R. M. Nucleic Acids Res 1990, 18, 6097. 37. Liu, J.; Louie, S.; Hsu, W.; Yu, K. M.; Nicholas, H. B., Jr.;
Rosen-quist, G. L. Am J Respir Cell Mol Biol 2008, 38, 738.
38. Huang, H. D.; Lee, T. Y.; Tzeng, S. W.; Wu, L. C.; Horng, J. T.; Tsou, A. P.; Huang, K. T. J Comput Chem 2005, 26, 1032. 39. Hunter, T. Philos Trans R Soc Lond B Biol Sci 1998, 353, 583. 40. Lee, T. Y.; Huang, H. D.; Hung, J. H.; Huang, H. Y.; Yang, Y. S.;
