行政院國家科學委員會專題研究計畫 成果報告
蛋白質立體結構預測（3/3）
計畫類別： 個別型計畫 計畫編號： NSC93-3112-B-002-022- 執行期間： 93 年 05 月 01 日至 94 年 04 月 30 日 執行單位： 國立臺灣大學資訊工程學系暨研究所 計畫主持人： 高成炎 共同主持人： 楊進木 計畫參與人員： 詹鎮熊 博士 報告類型： 完整報告 報告附件： 出席國際會議研究心得報告及發表論文 處理方式： 本計畫可公開查詢中 華 民 國 94 年 8 月 8 日
行政院國家科學委員會補助專題研究計畫
■ 成 果 報 告
□期中進度報告
蛋白質立體結構預測（3/3）
計畫類別：■ 個別型計畫 □ 整合型計畫
計畫編號：NSC 93－3112－B－002－022－
執行期間：
93 年 5 月 1 日至 94 年 4 月 30 日
計畫主持人：高成炎 教授
共同主持人：張文章、楊進木、林智仁 教授
計畫參與人員：詹鎮熊 博士
成果報告類型(依經費核定清單規定繳交)：□精簡報告 ■完整報告
本成果報告包括以下應繳交之附件：
□赴國外出差或研習心得報告一份
□赴大陸地區出差或研習心得報告一份
□出席國際學術會議心得報告及發表之論文各一份
□國際合作研究計畫國外研究報告書一份
處理方式：除產學合作研究計畫、提升產業技術及人才培育研究計畫、
列管計畫及下列情形者外，得立即公開查詢
□涉及專利或其他智慧財產權，□一年□二年後可公開查詢
執行單位：國立臺灣大學資訊工程學系暨研究所
中 華 民 國 94 年 7 月 30 日
The results accomplished in this project has been published in Bioinformatics,
Vol. 21, pp. 1415-1420, 2005. (SCI, 2004 impact factor: 6.7) The manuscript
is attached in this report.
Cysteine separations profiles (CSP) on protein sequences infer disulfide
connectivity
East Zhao1, Hsuan-Liang Liu2, Chi-Hung Tsai1, Huai-Kuang Tsai1, Chen-hsiung Chan1, Cheng-Yan Kao1*
1Bioinformatics Laboratory, Department of Computer Science and Information Engineering,
National Taiwan University
No. 1 Sec. 4 Roosevelt Rd., Taipei, Taiwan 106
2Department of Chemical Engineering and Graduate Institute of Biotechnology, National Taipei
University of Technology
No. 1 Sec. 3 Chung-Hsiao E. Rd., Taipei, Taiwan 10608
*Corresponding author: Ph: +886-2-23625336 ext. 401 Fax: +886-2-23658741
Abstract
Motivation: Disulfide bonds play an important role in protein folding. A precise prediction of disulfide connectivity can strongly reduce the conformational search space and increase the
accuracy in protein structure prediction. Conventional disulfide connectivity predictions use
sequence information, and prediction accuracy is limited. Here, by using an alternative scheme
with global information for disulfide connectivity prediction, higher performance is obtained with
respect to other approaches.
Result: Cysteine separation profiles (CSPs) have been used to predict the disulfide connectivity of proteins. The separations among oxidized cysteine residues on a protein sequence have been
encoded into vectors named cysteine separation profiles (CSPs). Through comparisons of their
CSPs, the disulfide connectivity of a test protein is inferred from a non-redundant template set.
For non-redundant proteins in SwissProt 39 (SP39) sharing less than 30% sequence identity, the
prediction accuracy of a four-fold cross validation is 49%. The prediction accuracy of disulfide
connectivity for proteins in SwissProt 43 (SP43) is even higher (53%). The relationship between
the similarity of CSPs and the prediction accuracy is also discussed. The method proposed in this
work is relatively simple and can generate higher accuracies comparing to conventional methods.
It may be also combined with other algorithms for further improvements in protein structure
prediction.
Availability: The program and datasets are available from the authors upon request. Contact: cykao@csie.ntu.edu.tw
1. Introduction
A disulfide bond is a strong covalent bond between two cysteine residues in proteins. It
plays a key role in protein folding and in determining the structure/function relationships of
proteins (Abkevich and Shakhnovich, 2000; Wedemeyer et al., 2000; Welker et al., 2001). In
addition, it is important in maintaining a protein in its stable folded state. A disulfide
connectivity pattern can be used to discriminate the structural similarity between proteins
(Chuang et al., 2003). In protein folding prediction, the knowledge of the locations of
disulfide bonds can dramatically reduce the search in conformational space (Skolnick et al.,
1997; Huang et al., 1999). Therefore, a higher performance in predicting disulfide
connectivity pattern is likely to increase the accuracy in predicting the three-dimensional
(3D) structures of proteins through the reduction of the number of steps during
conformational space search.
Generally, the prediction of disulfide connectivity pattern in proteins consists of two
consecutive steps. Firstly, the disulfide bonding state of each cysteine residue in a protein is
predicted based on its amino acid sequence and evolutionary information using various
algorithms, such as neural networks (Fariselli et al., 1999; Fiser and Simon, 2000), support
vector machines (Chen et al., 2004), and hidden Markov models (Martelli et al., 2002).
Secondly, the location of disulfide bonds is subsequently predicted based on the bonding
state of each cysteine residue using algorithms such as Monte Carlo (MC) simulated
annealing together with weighted graph matching (Fariselli and Casadio, 2001) and
prediction accuracy of the oxidation state of cysteine residues has reached 90% (Chen et al.,
2004) and can be used confidently. However, the task of predicting disulfide connectivity
remains challenging. The best prediction accuracy ever reported so far is only 44% (Vullo
and Frasconi, 2004), in which recursive neural network was used to score connectivity
patterns represented in unidirected graphs. Such prediction accuracy is still far from being
usable, although it is much higher than that by a random predictor.
In this work, cysteine separation profiles (CSPs) of proteins are adopted for the
prediction of disulfide connectivity. It has been shown that proteins with similar disulfide
bonding patterns also share similar folds (Chuang et al., 2003; van Vlijmen et al., 2004).
Theoretical work has suggested that disulfide bonds may stabilize the structures of protein
fragments between the connected cysteine residues (Abkevich and Shakhnovich, 2000);
therefore, the separations between oxidized cysteine residues may be used in the task of
predicting disulfide connectivity. Previous works on disulfide connectivity predictions have
used graphs to represent disulfide connection patterns (Fariselli and Casadio, 2001; Vullo
and Frasconi, 2004). Protein sequences, contact potentials, and evolutionary information
have been well used to score various connection patterns. The present approach encodes
separations among cysteine residues into the form of vectors. The prediction of disulfide
connectivity is based on the comparisons of vectors from testing and template dataset, in
which similar vectors imply similar connection patterns. The method proposed here (CSP) is
2. System and Methods
2.1 Datasets
The datasets used to evaluate the predicting power of CSPs were constructed from
SwissProt release No. 39 (Bairoch and Apweiler, 2000), including sequences with annotated
disulfide bridges. Protein sequences in SwissProt release No. 39 are filtered according to
procedures described in two previous works (Fariselli and Casadio, 2001; Vullo and
Frasconi, 2004). This dataset is denoted as ‘SP39’. Another dataset based on SP39 was also
constructed; redundant sequences with pairwise sequence identity more than 30% were
removed. This non-redundant set is denoted as ‘SP39-ID30’. SP39-ID30 is used to
investigate the effects of sequence identities on the prediction accuracy of CSP.
Another dataset was further constructed to verify the predicting power of CSP. The
same filter procedures were applied to sequences in SwissProt release No. 43, where
sequences in release 39 were excluded. Thus it is possible to predict proteins newly added to
SwissProt database between releases No. 39 and No. 43. This set is denoted as ‘SP43’.
Redundant sequences with pairwise sequence identity more than 25% in SP43 were also
removed. The template set used to predict disulfide connectivity in SP43 was constructed
from SwissProt release 39. Sequences in this set were filtered as in SP39 and SP43, except
for the PDB filter. Only sequences sharing less than 30% identity with those in SP43 were
kept. This template set is denoted as ‘SP39-TEMPLATE’.
The numbers of sequences divided according to the number of disulfide bridges in
2.2 Basic Assumption
Similar disulfide bonding patterns infer similar protein structures regardless of sequence
identity (Chuang et al., 2003). Figure 1 shows an example of two proteins with the same disulfide
bonding patterns. Tick anticoagulant peptide (serine protease inhibitor, PDB id 1TAP) (Antuch et
al., 1994) and cacicludine (calcium channel blocker, PDB id 1BF0) (Gilquin et al., 1999) exhibit
the same disulfide connectivity [1-6, 2-3, 4-5], which means that the first oxidized cysteine is
connected with the sixth one, the second with the third, and the forth with the fifth. These two
proteins share sequence identity of only 18.2%, but with a Cα root-mean-square deviation
(RMSD) of 3.6Å (Chuang et al., 2003). Although the sequence identity is below the twilight zone,
the structure and separations among cysteine residues are similar for these two proteins. The
residue numbers for cysteines in the two proteins are [5, 15, 33, 39, 55, 59] and [7, 16, 32, 40, 53,
57], respectively. The positions and separations of cysteine residues are similar for these two
proteins. It is likely that cysteine separations are related to disulfide connectivity patterns, and
through the comparison of cysteine separation profiles (CSPs), the disulfide connectivity patterns
may be inferred and predicted.
2.3 Cysteine Separation Profile (CSP) and evaluation of prediction accuracy
Cysteine separation profiles (CSPs) contain cysteine separation information. Protein x
with n disulfide bonds and 2n cysteine residues has a cysteine separation profile (CSPx)
defined as
)
(
1, 2,..., 2 −1) (
= 2 − 1, 3− 2,..., 2 − 2 −1 = n n n x C C C C C C s s s CSPseparation between cysteines Ci and Ci+1. By this definition, a protein with disulfide bonds
will have a CSP.
The divergence, D, between two CSPs is defined as follows:
∑
− = i Y i X i s s Dwhere X and are the ith separations for CSPs of two different proteins X and Y.
i
s siY
The CSP of a test protein was then compared with all CSPs of template proteins. The
disulfide connectivity pattern of the test protein can be predicted as that of the template
protein with the most similar CSP, i.e., with the smallest divergence value, D. If the
divergence D between two CSPs equals to 0, the CSPs was termed ‘matched profiles’,
otherwise they are ‘mismatched profiles’. If more than one template proteins are matched,
one of the templates is randomly selected for the prediction. The ambiguous situations are
rare, only less than 2% are observed.
Prediction accuracy of our method was evaluated with Qp and Qc value, which are the
fraction of proteins with correct disulfide connectivity prediction and is defined as:
, p c p c p c C C Q Q T T = =
where Cp is number of proteins with all the disulfide connectivity correctly predicted;
Tp is the total number of test proteins; Cc is the number of disulfide connectivity correctly
3. Results
3.1 Four-fold cross validation
In order to compare with other approaches for disulfide connectivity prediction, similar criteria
were used to select our dataset. The same four-fold cross validation has been applied to our
datasets. The SP39 and SP39-ID30 datasets were divided into 4 subsets, and the disulfide
connectivity prediction was repeated 4 times. For each prediction, one of the 4 subsets was used
as the test set and the other 3 subsets were put together to form a template set. The final
prediction accuracy was averaged over the four prediction results.
Table 2 summarizes the disulfide connectivity prediction results obtained from this
study as well as those obtained from the previous works (Fariselli and Casadio, 2001; Vullo
and Frasconi, 2004). “Frequency” is a trivial method, where the prediction is based on most
frequently observed pattern in the training set. ‘MC graph-matching’ and ‘NN
graph-matching’ are both based on a graph representation of disulfide bonding patterns,
using Monte Carlo and Neural Networks for pattern recognition, respectively (Fariselli and
Casadio, 2001). The results termed BiRnn are obtained from recursive neural networks with
sequence and evolutionary information (Vullo and Frasconi, 2004), the disulfide
connectivity patterns are also represented using graphs. The prediction results from this
work are termed CSP, with dataset noted in the parenthesis. The prediction results are
divided according to the number of disulfide bridges.
The average value of Qp using CSP is 0.81 for SP39. However, redundant sequences
mismatched profiles patterns. The number of matched profile patterns is high, and is likely
to be resulted from redundant and homologous sequences in SP39 dataset. The redundancy
may have caused over fitting in SP39, even with four-fold cross validation. In order to
control and test over fitting, we extracted the sequences with pairwise sequence identities
less than 30% from SP39 and then generated another dataset, SP39-ID30. The average value
of Qp (B = 2~5) using CSP is 49% for SP39-ID30. With redundant sequences removed, the
four-fold cross validation prediction accuracy of CSPs is higher than the best results ever
reported from previous works.
The prediction accuracies for protein chains with different disulfide bridge numbers are
all significantly higher for ‘CSP (SP39)’. For proteins with 2, 4, and 5 disulfide bridges, the
prediction accuracies in ‘CSP (SP39-ID30)’ are higher than other works. The prediction
accuracy for proteins with 3 disulfide bridges is 2% lower than that of ‘BiRnn-1 profile’, but
is still significantly higher than those from other works. 3.2 Handout Prediction of New Sequences from SP43
We further validate CSP on a new dataset, SP43, which contains new sequences not
seen in SwissProt release 39. We use SP39-TEMPLATE as template set to predict disulfide
connectivity patterns of new sequences in SP43. The pairwise identities of sequences in the
template set and SP43 are less than 30%, with template sequences sharing higher identities
with those in SP43 being removed. The overall prediction accuracy in SP43 dataset is 53%,
which shows significant improvement over the prediction on the other dataset, SP39. The
bridges, the prediction accuracies in SP43 dataset are higher than those obtained with
four-fold cross validation in SP39-ID30 dataset. This implies that increasing even the
number of non-redundant templates may improve the prediction accuracy of CSP. 3.3 Examples
Three examples of CSP matching are listed in Table 3. These examples are taken from
the SSDB database (Chuang et al., 2003). The CSPs for template and query protein
sequences, as well as their divergence score D, disulfide connectivity patterns, and sequence
identities, are shown in Table 3. In the three examples, the divergence scores are all smaller
than 10, implying that they share similar disulfide positioning and connectivity patterns. The
sequence identities in the three examples are all lower than 20%, thus structure similarity
from sequence homology can be ruled out.
The structures and sequences of these examples are illustrated in Figures 1-3. The first
example is shown in Figure 1. Tick anticoagulant peptide (serine protease inhibitor, PDB id
1TAP) (Antuch et al., 1994) and cacicludine (calcium channel blocker, PDB id 1BF0)
(Gilquin et al., 1999) have a divergence score D=8, their disulfide connectivity pattern is
[1-6, 2-3, 4-5]. Example 2 is illustrated in Figure 2. Thionin (toxic arthropod protein, PDB
id 1GPS) (Bruix et al., 1993) and brazzein (thermostable sweet-tasting protein, PDB id 1brz)
(Caldwell et al., 1998) share 18.8% sequence identity. Their divergence score D is 6, and the
disulfide connectivity pattern is [1-8, 2-5, 3-4, 6-7]. The third example (Figure 3), C-type
lectin carbohydrate recognition domain of human tetranectin (PDB id 1TN3) (Kastrup et al.,
also have a divergence score of D=6 . Their sequence identity is 17.7% and the
connectivity pattern is [1-2, 3-6, 4-5]. For all proteins, the oxidized cysteine residues are
indicated in black. Cysteine residues on sequences are highlighted in bold and underline. In
each case, the cysteine residues are positioned in similar sites along the sequence, and the
separations among these cysteine residues are nearly identical.
4. Discussions
The number of possible disulfide connectivity patterns increases rapidly with number
of disulfide bridges. For a protein with n disulfide bridges (n*2 oxidized cysteines), the
number of possible disulfide connectivity patterns Np can be formulated as follows:
∏
≤ − = − = ⎟⎟ ⎠ ⎞ ⎜⎜ ⎝ ⎛ ⎟⎟ ⎠ ⎞ ⎜⎜ ⎝ ⎛ − ⎟⎟ ⎠ ⎞ ⎜⎜ ⎝ ⎛ − ⎟⎟ ⎠ ⎞ ⎜⎜ ⎝ ⎛ = n i p n i n n n n N (2 1)!! (2 1) ! 2 2 2 4 2 2 2 2 2 2 L .Table 4 lists the number of possible disulfide connectivity patterns for proteins with
different disulfide bridge numbers. The use of CSPs may be obscure at first, since the
rapidly increasing number of pattern cannot be covered exhaustively. However, the observed
numbers of patterns in PDB peak at 5 disulfide bridges, and decline afterward. Only 45
patterns are observed for protein chains with 5 disulfide bridges, as opposed to possible 945
patterns expected. These results imply that the disulfide connectivity pattern of a protein
sequence can be predicted from a limited set of templates.
One limitation of our approach is that a pattern not presented in the training set cannot
be predicted correctly. Other machine learning approach has to enumerate all possible
patterns to obtain a prediction with maximum score (Vullo and Frasconi, 2004); therefore it
of all possible patterns is expensive (Vullo and Frasconi, 2004), our approach can achieve
comparable prediction performance in a much simpler and faster algorithm.
The use of divergence score D for profile matching deserves more discussion, e.g.
what is the divergence score cutoff for a reliable prediction? The prediction accuracies for
protein chains with different divergence score coverage are shown in Figure 4. The
divergence coverage means that a profile matches with divergence score smaller than or
equal to that specified. For example, coverage 5 means profiles matched with a divergence
score equal to or smaller than 5. Prediction results of the three datasets are illustrated. As
can be seen, when divergence coverage is 0, which means the profiles are ‘matched profiles’,
the prediction accuracy is 100%. The prediction accuracies become lower as divergence
coverage increases. With divergence coverage of 50, the prediction accuracy is slightly
higher than the overall accuracy. Divergence coverage can be used as a decision value for
adoption of CSP or machine-learning approaches. However, these divergence scores are not
normalized according to the number of disulfide bridges. Sequences with low divergence
matches can be predicted with CSP, and the connectivity patterns of other sequences can be
elucidated with neural network (Vullo and Frasconi, 2004), support vector machine (Chen et
al., 2004), or other machine-learning approaches.
5. Conclusions
In this work, we have shown that cysteine separation profiles (CSPs) can be used in
predicting disulfide connectivity patterns based on the hypothesis that proteins with similar
accuracy of CSP proposed in this study is higher than those obtained by other approaches. The
handout prediction of new sequences in SP43 dataset can reach 53%. The method mentioned here
is extremely simple; therefore the computation time is minimum comparing to other methods.
The rational behind our method is completely different from previous studies using sequence and
evolutionary information. Our method suggests that topology itself may be an important factor in
disulfide connectivity, as it has been proposed by theoretical study (Abkevich and Shakhnovich,
2000) and observations in structure databases (Chuang et al., 2003). Although many efforts have
been made to predict the disulfide connectivity patterns, current prediction accuracy is limited
around 50%. However, by combining CSP and other algorithms proposed previously (Fariselli
and Casadio, 2001; Vullo and Frasconi, 2004), it is possible to further improve the prediction
accuracy. The use of predicted disulfide connectivity patterns in ab initio protein structure
prediction and other applications would become more reliable in the foreseeable future.
6. Acknowledgements
The authors thank National Science Council of Taiwan for financial support (project number
NSC-93-3112-B-002-022).
7. References
Abkevich, V. I. and Shakhnovich, E. I. (2000) What can disulfide bonds tell us about protein energetics, function and folding: Simulations and bioinformatics analysis. J. Mol. Biol., 300, 975-985.
NMR solution structure of the recombinant tick anticoagulant protein (rtap), a factor Xa inhibitor from the tick Ornithodoros moubata. FEBS Lett., 352, 251-257.
Bairoch, A. and Apweiler, R. (2000) The Swiss-Prot protein sequence database and its supplement TrEMBL in 2000. Nucleic Acids Res., 28, 45-48.
Bruix, M., Jimenez, M. A., Santoro, J., Gonzalez, C., Colilla, F. J., Mendez, E. and Rico, M. (1993) Solution structure of gamma 1-H and gamma 1-P thionins from barley and wheat endosperm determined by 1H-NMR: A structural motif common to toxic arthropod proteins. Biochemistry, 32, 715-724.
Caldwell, J. E., Abildgaard, F., Dzakula, Z., Ming, D., Hellekant, G. and Markley, J. L. (1998) Solution structure of the thermostable sweet-tasting protein brazzein. Nat. Struct. Biol., 5, 427-431.
Chen, Y.-C., Lin, Y.-S., Lin, C.-J. and Hwang, J.-K. (2004) Prediction of the bonding states of cysteines using the support vector machines based on multiple feature vectors and cysteine state sequences. Proteins., 55, 1036-1042.
Chuang, C.-C., Chen, C.-Y., Yang, J.-M., Lyu, P.-C. and Hwang, J.-K. (2003) Relationship between protein structures and disulfide-bonding patterns. Proteins., 55, 1-5.
Fariselli, P., Martelli, P.L., & Casadio, R. (2002). A neural network base method for prediction the disulfide connectivity in proteins. In Damiani, E., et al. (Eds.), Knowledge based
intelligent information engineering systems and allied technologies (KES 2002), Vol. 1,
pp. 464-468. IOS Press.
Fariselli, P. and Casadio, R. (2001) Prediction of disulfide connectivity in proteins.
Bioinformatics, 17, 957-964.
Fariselli, P., Riccobelli, P. and Casadio, R. (1999) Role of evolutionary information in predicting the disulfide-bonding state of cysteine in proteins. Proteins., 36, 340-346.
Fiser, A. and Simon, I. (2000) Predicting the oxidation state of cysteines by multiple sequence alignment. Bioinformatics, 16, 251-256.
Fukuda, K., Mizuno, H., Atoda, H. and Morita, T. (2000) Crystal structure of flavocetin-a, a platelet glycoprotein Ib-binding protein, reveals a novel cyclic tetramer of c-type
lectin-like heterodimers. Biochemistry, 39, 1915-1923.
Gilquin, B., Lecoq, A., Desne, F., Guenneugues, M., Zinn-Justin, S. and Menez, A. (1999) Conformational and functional variability supported by the BPTI fold: Solution structure of the Ca2+ channel blocker calcicludine. Proteins., 34, 520-532.
Huang, E. S., Samudrala, R. and Ponder, J. W. (1999) Ab initio fold prediction of small helical proteins using distance geometry and knowledge-based scoring functions. J. Mol. Biol., 290, 267-281.
Kastrup, J. S., Nielsen, B. B., Rasmussen, H., Holtet, T. L., Graversen, J. H., Etzerodt, M., Thogersen, H. C. and Larsen, I. K. (1998) Structure of the c-type lectin carbohydrate recognition domain of human tetranectin. Acta Crystallogr. D Biol. Crystallogr., 54, 757-766.
Martelli, P. L., Fariselli, P., Malaguti, L. and Casadio, R. (2002) Prediction of the
disulfide-bonding state of cysteines in proteins at 88% accuracy. Protein Sci., 11, 2735-2739.
Skolnick, J., Kolinski, A. and Ortiz, A. R. (1997) MONSSTER: A method for folding globular proteins with a small number of distance restraints. J. Mol. Biol., 265, 217-241.
van Vlijmen, H. W. T., Gupta, A., Narasimhan, L. S. and Singh, J. (2004) A novel database of disulfide patterns and its application to the discovery of distantly related homologs. J. Mol.
Biol., 335, 1083-1092.
Vullo, A. and Frasconi, P. (2004) Disulfide connectivity prediction using recursive neural networks and evolutionary information. Bioinformatics, 20, 653-659.
Wedemeyer, W. J., Welker, E., Narayan, M. and Scheraga, H. A. (2000) Disulfide bonds and protein folding. Biochemistry, 39, 4207-4216.
Welker, E., Wedemeyer, W. J., Narayan, M. and Scheraga, H. A. (2001) Coupling of
conformational folding and disulfide-bond reactions in oxidative folding of proteins.
Table 1
Number of chains in the datasets, divided according to the number of disulfide bridges (B).
Datasets B = 2 B = 3 B = 4 B = 5 B = 2~5
SP39 150 149 105 45 449
SP39-ID30 92 81 43 28 244
SP39-TEMPLATE 244 198 98 45 585
Table 2
Comparison among different disulfide connectivity prediction algorithms.
B = 2 B = 3 B = 4 B = 5 B = 2~5 Algorithms Qp Qc Qp Qc Qp Qc Qp Qc Qp Qc Frequencya 58% 58% 29% 37% 1% 10% 0% 23% 29% 32% MC graph-matchingb 56% 56% 21% 36% 17% 37% 2% 21% 29% 38% NN graph-matchingc 68% 68% 22% 37% 20% 37% 2% 26% 34% 42% BiRnn-1 sequenced 59% 59% 17% 30% 10% 22% 4% 18% 28% 32% BiRnn-1 profiled 65% 65% 46% 56% 24% 32% 8% 27% 42% 46% BiRnn-2 sequenced 59% 59% 22% 34% 18% 30% 8% 24% 31% 37% BiRnn-2 profiled 73% 73% 41% 51% 24% 37% 13% 30% 44% 49% CSP (SP39)e 89% 89% 81% 84% 81% 85% 51% 60% 81% 81% CSP (SP39-ID30)f 74% 74% 44% 53% 26% 44% 18% 31% 49% 52% CSP (SP43)g 71% 71% 49% 58% 30% 40% 28% 33% 53% 53%
a Prediction accuracy reported by Vullo and Frasconi (2004).
b Prediction accuracy reported by Fariselli and Casadio (2001).
c Prediction accuracy reported by Fariselli et al. (2002).
d Prediction accuracy reported by Vullo and Frasconi (2004).
e Prediction accuracy of CSP on SP39 with redundant sequences retained.
f Prediction accuracy of CSP with redundant sequences removed.
Table 3
Examples of cysteine separation profiles (CSP).
Template PDB id Template CSP Query PDB id Query CSP Disulfide Connectivity Pattern Divergence (D) Sequence Identity 1TAP (10, 18, 6, 16, 4) 1BF0 (9, 16, 8, 13, 4) [1-6, 2-3, 4-5] 8 18.2% 1GPS (11, 6, 4, 10, 7, 2, 4) 1BRZ (12, 6, 4, 11, 10, 2, 3) [1-8, 2-5, 3-4, 6-7] 6 18.8% 1TN3 (10, 17, 75, 16, 8) 1C3A:A (11, 17, 72, 17, 8) [1-2, 3-6, 4-5] 6 17.7%
Table 4
Number of possible disulfide connectivity patterns (Np) for protein chains with different disulfide
bridge numbers. Number of disulfide bridges (B) Np a Observed N pb B = 2 3 3 B = 3 15 15 B = 4 105 43 B = 5 945 45 B = 6 10395 29 B = 7 135135 14
a Number of possible disulfide connectivity patterns.
b Observed number of disulfide connectivity patterns. Statistics obtained from the SSDB database
Figure legends
Figure 1
The structures of two proteins with low sequence identity but share the same disulfide bonding
patterns. (a) anticoagulant protein (PDB id 1TAP), (b) calcium channel blocker (PDB id 1BF0),
and (c) the sequences of the two protein, with cysteine residues highlighted with bold and
underline. Both proteins have three disulfide bonds ([1-6, 2-3, 4-5]) and BPTI-like structures; the
sequence identity is 18.2%.
Figure 2
The structures of (a) thionin, toxic arthropod protein (PDB id 1GPS), (b) brazzein, thermostable
sweet-tasting protein (PDB id 1brz), and (c) their sequences with cysteine highlighted. The
divergence score D between these two protein sequences is 6. Both proteins have disulfide
connectivity [1-8, 2-5, 3-4, 6-7] and their sequence identity is 18.8%.
Figure 3
Structures of (a) C-type lectin carbohydrate recognition domain of human tetranectin (PDB id
1TN3), (b) flavocetin-A from Habu snake venom (PDB id 1C3A:A), and (c) their sequences with
cysteine residues highlighted. The divergence score D between these two protein sequences is 6.
Both proteins have disulfide connectivity [1-2, 3-6, 4-5] and their sequence identity is 17.7%.
Figure 4
Prediction accuracy of the datasets with various divergence coverage, which means that a profile
(Qp) is higher with lower divergence coverage. With D≤50 the prediction accuracy is still
slightly higher than the overall Qp. See text for details.
