Open Access
Methodology article
(PS)
2
-v2: template-based protein structure prediction server
Chih-Chieh Chen
1, Jenn-Kang Hwang
1,2,3and Jinn-Moon Yang*
1,2,3Address: 1Institute of Bioinformatics, National Chiao Tung University, Hsinchu 30050, Taiwan, Republic of China, 2Department of Biological
Science and Technology, National Chiao Tung University, Hsinchu 30050, Taiwan, Republic of China and 3Molecular Bioinformatics Center,
National Chiao Tung University, Hsinchu 30050, Taiwan, Republic of China
Email: Chih-Chieh Chen - [email protected]; Jenn-Kang Hwang - [email protected]; Jinn-Moon Yang* - [email protected]
* Corresponding author
Abstract
Background: Template selection and target-template alignment are critical steps for template-based modeling (TBM) methods. To identify the template for the twilight zone of 15~25% sequence similarity between targets and templates is still difficulty for template-based protein structure prediction. This study presents the (PS)2-v2 server, based on our original server with numerous
enhancements and modifications, to improve reliability and applicability.
Results: To detect homologous proteins with remote similarity, the (PS)2-v2 server utilizes the
S2A2 matrix, which is a 60 × 60 substitution matrix using the secondary structure propensities of 20 amino acids, and the position-specific sequence profile (PSSM) generated by PSI-BLAST. In addition, our server uses multiple templates and multiple models to build and assess models. Our method was evaluated on the Lindahl benchmark for fold recognition and ProSup benchmark for sequence alignment. Evaluation results indicated that our method outperforms sequence-profile approaches, and had comparable performance to that of structure-based methods on these benchmarks. Finally, we tested our method using the 154 TBM targets of the CASP8 (Critical Assessment of Techniques for Protein Structure Prediction) dataset. Experimental results show that (PS)2-v2 is ranked 6th among 72 severs and is faster than the top-rank five serves, which utilize
ab initio methods.
Conclusion: Experimental results demonstrate that (PS)2-v2 with the S2A2 matrix is useful for
template selections and target-template alignments by blending the amino acid and structural propensities. The multiple-template and multiple-model strategies are able to significantly improve the accuracies for target-template alignments in the twilight zone. We believe that this server is useful in structure prediction and modeling, especially in detecting homologous templates with sequence similarity in the twilight zone.
Background
For template-based modeling (TBM) and fold recognition methods, a prediction model can be built based on the coordinates of the appropriate template(s) [1]. These approaches generally involve four steps: 1) a
representa-tive protein structure database is searched to identify a template that is structurally similar to the protein target; 2) an alignment between the target and the template is generated that should align equivalent residues together as in the case of a structural alignment; 3) a prediction
Published: 31 October 2009
BMC Bioinformatics 2009, 10:366 doi:10.1186/1471-2105-10-366
Received: 29 June 2009 Accepted: 31 October 2009 This article is available from: http://www.biomedcentral.com/1471-2105/10/366
© 2009 Chen et al; licensee BioMed Central Ltd.
This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
structure of the target is built based on the alignment and the selected template structure, and 4) model quality eval-uation. The first two steps significantly affect the quality of the final model prediction in TBM methods.
The secondary structure of a protein is often more con-served than the amino acid sequence, and the prediction accuracy of the secondary structure has been achieved ~80% on average. Recently, a number of methods, inte-grating secondary structures (i.e., α-helix, β-strand and coil) with primary amino acid sequences, have success-fully detected the homologs with remote similarity for automated comparative modeling [2-6] and fold recogni-tion [7-12]. These methods often used two separated sub-stitution matrices [9,10,13] to score secondary structures and primary amino acids, respectively, for aligning a resi-due pair. The separated matrices are unable to reflect the real score because the amino acid type often prefers to a specific secondary structure.
Here, we have developed a substitution matrix, called S2A2, which considers the properties of the secondary structures and amino acid types. The S2A2 is a 60 × 60 matrix that considers all possible pair combination of 20 amino acid types and three secondary structure elements. This matrix was evaluated on the Lindahl benchmark [14] for fold recognition and the ProSup benchmark [15] for alignment accuracies. According to these evaluation results, the S2A2 matrix has higher accuracy than position specific scoring matrix (PSSM) generated by PSI-BLAST and prof_sim for fold recognition and sequence align-ments. By integrating the S2A2 matrix and PSSM, each having a unique scoring mechanism, the (PS)2-v2 server
blends the sequence profile and secondary structure infor-mation so that they work cooperatively.
Numerous enhancements and modifications were applied to original (PS)2 servers (namely (PS)2-original)
[16] and (PS)2-CASP8 [17] which participates the CASP8
experiment, thereby improving the reliability and applica-bility of the method. There are four main differences in methodology between the present server ((PS)2-v2) and
our previous works (Table 1). First, (PS)2-v2 integrates
S2A2 matrix and PSSM for the template selection and the target-template alignment to replace a consensus strategy applied in the (PS)2-original server. Second, we modified
the SSEARCH [18] search method to replace the PSI-BLAST search method and Smith-Waterman algorithm applied in the (PS)2-original server and (PS)2-CASP8,
respectively. Third, (PS)2-v2 utilized a new multiple
tem-plate method for modeling different domains of the target sequence. Finally, (PS)2-v2 added a multiple model
strat-egy and utilized ProQ [19] to assess and select the final model. We have assessed the prediction accuracy of the (PS)2-v2 server based on the 154 TBM targets of the
CASP8 dataset. The experimental results show that the S2A2 matrix, multiple-template and multiple-model strat-egies are able to significantly improve the accuracies for protein structure prediction and modeling when the sequence similarity between the template and the target is in the twilight zone.
Methods
Figures 1 and 2 show the framework of the (PS)2-v2 server
for protein structure prediction. (PS)2-v2 uses the S2A2
matrix and the PSSM for the template selection and the target-template alignment. (PS)2-v2 first applied the query
sequence to generate a PSSM by running three iterations of PSI-BLAST against a non-redundant sequence UniRef90 [20] with an E-value cutoff of 0.001. The PSSM was then used as the input for the PSIPRED [21] tool to predict the secondary structure of this query. We then modified the SSEARCH [18] search method, using the S2A2 matrix and the PSSM as the scoring matrices, to identify the template(s) from the protein structure library, and to generate the target-template alignment(s). The
Table 1: The essential differences of (PS)2-original, (PS)2-CASP8 and (PS)2-v2
Steps (PS)2-original [16] (PS)2-CASP8 [17] (PS)2-v2
1. Template search Consensus of PSI-BLAST and IMPALA
S2A2+PSSM with a self-developed aligned tool using dynamic
programming
S2A2+PSSM with a modified SSEARCH program [18]
2. Target-template alignment Consensus of PSI-BLAST, IMPALA and T-coffee
S2A2+PSSM with a self-developed aligned tool using dynamic
programming
S2A2+PSSM with a modified SSEARCH program [18]
3. Template Single template Single template Multiple templates
4. Model building MODELLER with single model MODELLER with single model MODELLER with multiple models
library consists of 20,982 non-redundant structures (April, 2008) selected from protein data bank (PDB) [22]. The secondary structures of each structure in the library were assigned using DSSP [23]. Based on various target-template alignments of top-ranking 5 selected target-templates, (PS)2-v2 generates 30 protein structures using
MODEL-LER [24]. Finally, the program ProQ was used to evaluate these models and to select the final model for the target. The S2A2 matrix, the aligned method, the modeling proc-ess and the final model selection are described in the fol-lowing subsections. The components of the (PS)2-v2
server were built using C, Perl and PHP (Additional file 1).
S2A2 matrix
A substitution matrix is the key component of protein sequence alignment methods. We developed the S2A2 substitution matrix (Figure 3 and Figure S1 in Additional file 2) applying a general mathematical structure [25]. To calculate the S2A2, 674 structural pairs (1,348 proteins) [26], which are structurally similar and with low sequence identity, were selected from SCOP 1.65 [27] based on two criteria: 1) the root-mean-square deviation (rmsd) of a protein pair was be less than 3.5 Å, with more than 70% of aligned residues included in the rmsd calculation, and 2) the sequence identity of a pair is less than 40%. The selected protein pairs had an average sequence identity of 26%, an average rmsd of 2.3 Å and average aligned resi-dues of 90% (207,492 aligned resiresi-dues out of 230,915
residues). The program DSSP was used to assign the sec-ondary structure for each residue of these 674 structural pairs. The eight types of the secondary structure used in DSSP were reduced to three commonly accepted types (H (helix), E (strand) and C (coil)) according to the follow-ing scheme: (H, G, I) → H; (E, B) → E; (T, S, blank) → C. The 20 amino acid types and 3 secondary structure types were converted into 60 residue-structure (RS) types. The S2A2 matrix (60 × 60) reveals substitution prefer-ences between homologs with low sequence identity, and was developed in a similar way to BLOSUM62 [25] based on these 674 structural pairs. The entry (Sij), which is the substitution score for aligning a RS letter i, j pair (1 ≤ i, j ≤ 60), of the S2A2 matrix is defined as Sij = λlog2(qij/eij), where is a scale factor, and qij and eij are the observed and expected probabilities, respectively, of the occurrence of each i, j pair. The observed probability is given by , where fij is the total number of align-ing i, j pairs in these 207,492 RS letters. The factor eij = pipj
if i = j; otherwise, eij = 2pipj (if i ≠ j), where pi is the back-ground probability of occurrence of the letter i, and equals . The substitution score is greater than zero (Sij > 0) if the observed probability is greater than the
fij k fmk m m /
∑
=∑
= 1 1 60 qii qik k i +∑
60≠ / 2 The framework of the (PS)2-v2 server for protein structure predictionFigure 1
The framework of the (PS)2-v2 server for protein structure prediction.
! " # $ % & ' () % * + ,(- * # . / # $ (. % # $ 0 1 2 3 4 5 % * + ,(- * # . / # $ (. % # $ 6 7 8 9 : ; < = > ?@ + ,% # A B CD 6 7 0 1 2
expected probability. By contrast, Sij < 0 if qij <eij. The is optimized by the SALIGN set [28], and is set to 1.6 accord-ing to the performance and efficiency.
Scoring and alignment methods
We modified the SSEARCH program [18], which used a rigorous Smith-Waterman algorithm [29], to search for similarity between a query sequence and template sequences in a library. We optimized the score between the query and template(s) using both S2A2 and PSSM
matrices based on alignment accuracies on the SALIGN set. The score is given as
S S w i j w S A i j w i j i j stru S A S , , max ( , ) ( , ) ( = + + − − −1 1 2 2 2 2 2 1 AA query i j gap S A S A S A pss PSSM i j S w w i j g w g 2 1 2 2 2 2 1 2 2 ) ( , ) ( , ) ( ) , − + + − mm i j gap S A S A S A pssm S,− +w w ( , )i j g + −( w )g ⎧ ⎨ ⎪ ⎪ ⎪⎪ ⎩ ⎪ ⎪ ⎪ ⎪ 1 2 2 2 2 1 2 2 0 Overview of the (PS)2-v2 server
Figure 2
Overview of the (PS)2-v2 server. The protein sequence of telomere replication protein Est3 (UniProt Q03096) in
Saccha-romyces cerevisiae was used as the query. (A) Input format of the (PS)2-v2 server. (B) Search results of a query protein,
compris-ing target name, sequence, predicted secondary structure, the graph of the aligned regions and the hits list of the templates of the query. (C) The selected template, target-template alignment and prediction structure of Est3. (D) The visualization of the predicted structure for Est3. (E) The model quality assessment.
where i and j are RS letters on the query and the template, respectively; wstru(i, j) is a structure-dependent scoring
weight, and is set to 1.3, 1.7 and 0.8 for α-helix, β-strand and coil, respectively; wS2A2 (here, wS2A2 is set to 0.64) is
the weight of the S2A2 matrix; S2A2(i, j) and PSSMquery(i,
j) are the scores of S2A2 and PSSM matrices, respectively,
when the RS letter i is aligned to the RS letter j. In addition, we considered structure-dependent gap penalty. Here, wgap
is a structure-dependent gapping weight, set to 2.0 (α-helix), 2.0 (β-strand) and 0.15 (coil), respectively; gS2A2 is
the gap opening penalty (set to 7.2) and the gap extension (set to 1.2) for the S2A2 matrix. These weights were opti-mized based on the SALIGN set. gpssm refers to the PSSM,
where the gap opening penalty is 11 and gap extension is 1 according to the default parameters of PSI-BLAST.
Statistics and template selection
SSEARCH provides the statistical significance for library searches. The local sequence similarity score (S) follows the extreme value distribution, so that P(S > x) = 1 -
exp(-Kmn exp(- x)) where m, n are the lengths of the query and library sequence. The score shows that the average score for an unrelated library sequence increases with the loga-rithm of the length of the library sequence. SSEARCH uses simple linear regression against the log of the library sequence length to calculate a normalized "z-score" with mean 50, regardless of library sequence length, and vari-ance 10. These z-scores can then be used with the extreme value distribution and the Poisson distribution to calcu-late the number of library sequences to obtain a score (i.e.
E-value) greater than or equal to the score obtained in the
search. The top-ranking 5 templates with the lowest E-val-ues were considered as the templates if the E-valE-val-ues < 0.1. For each structure in the top-ranking 5 templates, The (PS)2-v2 server generated six alternative target-template
alignments by using different S2A2-matrix (wS2A2)
weights, including 0, 0.2, 0.4, 0.64, 0.8 and 1.0. Finally, we yielded 30 target-template alignments for a target pro-tein.
The S2A2 substitution matrix Figure 3
The S2A2 substitution matrix. The scores are high if the residue-structure (RS) letters with similar residue types and the same secondary structure are aligned (red blocks). When two identical RS letters (e.g. diagonal entries) are aligned, the substi-tution scores are very high. In contrast, the scores are low when helix letters are aligned with strand letters (blue blocks).
! " # $ %
& & % $ # " ! & % $ # " !
' () * + , -. ( / 0 1 2 3 4 5 -6 ( 7 8 9 : ; < -G P H R K Q E D N T S A C M V I L F Y W G P H R K Q E D N T S A C M V I L F Y W G P H R K Q E D N T S A C M V I L F Y W
Model building and evaluation
Protein structure models were built using the homology modeling tool, MODELLER [24] according to the selected template(s) and target-template alignment(s) and then the ability to discriminate a correct protein model from incorrect models is critical when a server used multiple model methods. Here, we utilized the program ProQ [19] to assess the quality of protein models based on the LGscore [30] and a model was considered correct if the LGscore was greater than 1.5 [19]. The (PS)2-v2 server first
selected the protein model, generated by the first rank template with wS2A2 = 0.64 as the seed model. The LGscore
of the seed model was then compared with those of the other models based on the top-rank 5 templates with dif-ferent wS2A2 weights. A model was chosen as the final one
if it had the highest LGscore and its LGscore (> 0.7) was significantly better than that of the seed model. Other-wise, the server selected the seed model as the final model.
Multiple-template method
(PS)2-v2 considered a target as a multiple domain protein
if any region with >40 residues has non-aligned residues to the template(s) when using above "model building and evaluation" steps. For a multiple domain protein, (PS)2
-v2 automatically decided domain boundaries based on the borders of the large gaps between the target and the template(s), and repeatedly executed above steps to model the structures of the non-aligned residues (Figure 1). Finally, these multiple models were then used as struc-ture templates to generate the full-length final model for the query protein.
Utility
Input format
The (PS)2-v2 server is an easy-to-use web server (Figure 2).
Users input the query protein sequence in FASTA format. The server provides three modes (Automatic, Manual and 'Use this template') for choosing template(s) (Figure 2A). The default mode is 'Automatic'. In this mode, (PS)2-v2
automatically selects the modeling template(s). For the 'Manual' mode, our server enables users to assign specific template(s) from a list of candidates (Figure 2B). The 'Use this template' mode allows users to assign a specific pro-tein structure as the template. Finally, (PS)2-v2 transmits
the predicted results to the users by email addresses.
Output format
The (PS)2-v2 server typically yields a predicted structure
within 7 minutes if the query sequence length is ~200. The server shows a list of templates, selected template(s), tar-get-template alignment(s), predicted structure(s) and structure evaluations (Figures 2B and 2C). The predicted structures are visualized in PNG format generated by the MolScript [31] and Raster3D [32] packages. If the user clicks a PNG picture, then the corresponding protein 3D structure is also displayed on the AstexViewer [33] (Figure
2D). A user can download the predicted structure coordi-nates in the PDB format. The server also provides the tar-get-template alignments and the structure quality factors (Figure 2E).
Modeling of ever shorter telomeres 3
The ever shorter telomeres 3 (Est3, UniProt Q03096), which is essential for telomere replication in vivo, is a small regulatory subunit of telomerase from
Saccharomy-ces cerevisiae. According to structure prediction combined
with in vivo characterization, it has been reported that Est3 consists of a predicted OB-fold (oligosaccharide/oligonu-cleotide binding) with structurally similar to the OB-fold of the human Tpp1 protein [34]. Because of the limited degree of conservation between these two protein fami-lies, these two proteins could not be recognized from sim-ple sequence profile methods. Additionally, the original (PS)2 -v2 server could not recognize them.
For the target Est3, the (PS)2 -v2 server selected the
OB-fold domain of the Tpp1 protein (PDB code 2i46) from
Homo sapiens as the template [35], with an E-value of
0.014. This template shared only 17.6% sequence identity with the query sequence. Figure 2C shows the target-tem-plate alignment. The server successfully recognized Tpp1 as the template since the secondary structure identity between the template and Est3 was 66.7%. Our method could align together three conserved residues (i.e. Trp21/ Trp98, Asp86/Asp148 and Leu155/Leu204, in Est3 versus Tpp1; green blocks in Figure 2C), which are primarily involved in protein folding and/or stability of the OB-fold. Seven amino acid positions (yellow blocks in Figure 2C), which are structurally similar between the two pro-tein families, were also aligned. These 10 aligned residues, depicted in cyan, are clustered in the interior of the core of the OB-fold (Figure 2D).
Results and Discussion
In the template-based protein structure prediction, the template selection and the target-template alignment are the two critical steps, since they will significantly affect the quality of the final model prediction. The template selec-tions and the sequence alignments of the proposed method with the S2A2 matrix were evaluated by the Lin-dahl benchmark [14] and ProSup benchmark [15], respectively. In general, it is neither straightforward nor completely fair to compare the results of different fold-recognition and alignment methods given that each employs different sequence databases for sequence pro-files, structure databases for structure profiles and proper-ties, release dates, and scoring functions. Therefore, the comparisons between our methods and other published methods serve as an approximate guide. Here, we evalu-ated S2A2 matrix, PSI-BLAST and prof_sim using the same sequence database, UniRef90 [20], with the same param-eters to generate a PSSM for fold recognitions (Lindahl
benchmark) and sequence alignment (ProSup bench-mark). Furthermore, (PS)2-v2 was assessed and compared
with other 71 automatic servers on 154 TBM targets in CASP8. Please note that (PS)2-v2 did not participate in the
CASP8 experiment.
Evaluation of S2A2 matrix
The S2A2 matrix (60 × 60) offers insights about substitu-tion preferences of RS letters between homologous pro-tein sequences (Figure 3 and Figure S1 in Additional file 2). The highest substitution score in this matrix is for the alignment of a RS letter 'Wβ' with a RS letter 'Wβ', where
Wβ is the residue Trp with the β-strand structure (Figure S1
in Additional file 2). This substitution score is 6.2. In addition, the substitution scores are also high when two identical structural letters (e.g., diagonal entries) are aligned. For example, the alignment scores are 5.6 and 6.1 while 'Wα' and 'Cα' are aligned with 'Wα' and 'Cβ',
respec-tively; where Wα is the residue Trp with the α-helix
struc-ture and Cα represents the residue Cys with the α-helix structure. Most of the substitution scores are positive if two RS letters in the same secondary structure are aligned. On the other hand, the lowest substitution score is -7.8 in this S2A2. All of the substitution scores are low when the helix RS letters are aligned with the strand RS letters. The above relationships are in good agreement with biological functions of the relevant structures, showing that the matrix S2A2 embodies conventional knowledge about secondary structure conservation in proteins.
We compared the S2A2 matrix with BLOSUM62. The highest substitution scores are 6.2 (S2A2) and 11 (BLOSUM62). In contrast, the lowest score for S2A2 (-7.8) is much lower than that for BLOSUM62 (-4). The main reasons for this large difference are that α-helices and β-strands constitute very different protein secondary structures, and the RS letters pertaining to these two types of structure are more conserved than amino acid sequences. These results demonstrate that the RS letters with the S2A2 matrix may be able to more accurately find remote homologous sequences than simple amino acid sequence analyses.
Template selection
For the template selection, our method with S2A2 matrix was compared to other methods on Lindahl benchmark [14], which consists of 976 proteins, for the fold recogni-tion. This set included 555, 434 and 321 assignments for the family, superfamily and fold levels, respectively. The S2A2 matrix outperforms PSI-BLAST and is comparative to other methods on this set (Table 2). Our method (S2A2+PSSM), incorporating PSSM into S2A2, is the best for detecting similarity on the superfamily and fold levels for the top five ranks among the 10 comparative methods. At the superfamily level, the S2A2+PSSM, PSI-BLAST and
prof_sim [36] identified 75.6%, 49.1% and 61.3% of
Table 3: Comparing S2A2 matrix with other methods for sequence alignment accuracies on the ProSup benchmark
Method Tce T me Tie σ0e S2A2a 8732 947 7198 53.4 S2A2 + PSSMa 9470 868 6998 58.7 SSALNb 9256 1115 7245 58.3 SPARKSc - - - 57.2 prof_sim 8009 4505 3142 43.6 PSI-BLAST 6733 4938 3452 36.4 FASTAd 5340 3003 7452 31.4 aThis work.
bResults from Qiu and Elber [10]. cResults from Zhou and Zhou [9]. dResults from Domingues et al. [15]. eT
c and Tm are total numbers of correctly aligned and missed residue
pairs, respectively; Ti is the total number of incorrect aligned pairs; σ0 is the average percentage of correctly aligned residues
Table 2: Comparing S2A2 matrix with other methods for fold recognition on the Lindahl benchmark
Methods Family (%) Superfamily (%) Fold (%)
Top 1 Top 5 Top 1 Top 5 Top 1 Top 5
S2A2 a 77.1 85.1 43.8 63.1 26.5 50.8 S2A2+PSSM a 82.2 88.8 56.7 75.6 27.1 54.5 PSI-BLAST 74.4 79.5 38.5 49.1 4.4 14.6 prof_sim 80.7 86.5 50.9 61.3 22.1 39.6 RAPTORb 84.8 87.1 47.0 60.0 31.3 54.2 PROSPECT IIc 84.1 88.2 52.6 64.8 27.7 50.3 SPARKSd 81.6 88.1 52.5 69.1 24.3 47.7 FOLDproe 85.0 89.9 55.5 70.0 26.5 48.3 SP3 f 81.6 86.8 55.3 67.7 28.7 47.4 SP4 f 80.9 86.3 57.8 68.9 30.8 53.6 aThis work.
assignments, respectively. At the fold level, the S2A2+PSSM (54.5%) outperformed PSI-BLAST (14.6%) and prof_sim (39.6%) in identifying homologous pairs.
Target-template alignment
For the alignment between the target and the template, our algorithm was evaluated based on the ProSup bench-mark [15], which consists of 127 protein pairs with signif-icant structural similarity but with sequence identity of no more than 30%. The total numbers of correctly aligned residue pairs (Tc) of the S2A2, S2A2+PSSM, prof_sim and
SSALN [10] were 8732, 9470, 8009 and 9256 pairs, respectively (Table 3). The percentage σ0 (average percent-age of correctly aligned residues, divided by the length of the structural alignment per protein pair) of the S2A2, S2A2+PSSM, PSI-BLAST, prof_sim and SSALN were 53.4%, 58.7%, 36.4%, 43.6% and 58.3%, respectively. The S2A2 matrix is significantly better than those of sequence-based approaches, including FASTA, PSI-BLAST and prof_sim. The S2A2+PSSM achieved the highest alignment accuracy with slightly better than SPARKS [9] and SSALN, and much better than the other comparative methods.
CASP8 structure prediction
Our previous server ((PS)2-CASP8) and other 70 servers
participated in the CASP8 competition, involving 121 tar-gets for tertiary structure prediction. These 121 tartar-gets are officially classified into 154 TBM domains (Table S1 in Additional file 3). The accuracies of these 71 servers were evaluated based on the GDT_TS [37] scores directly
sum-marized from the CASP8 website http://prediction center.org/casp8/.
(PS)2-v2, (PS)2-original and (PS)2-CASP8 servers were
evaluated on these 154 TBM targets (Figure 4, Table 4 and Table S2 in Additional file 4). The sum of GDT_TS scores were 10331.4 ((PS)2-v2), 9954.4 ((PS)2-CASP8) and
9447.5 ((PS)2-original), respectively. (PS)2-v2 yielded 99
and 34 higher GDT_TS scores than (PS)2-original and
(PS)2-CASP8, respectively, among 154 targets. When the
sequence identity between the target and template was more than 30%, these three servers achieved similar GDT_TS scores. However, if the sequence identity was less than 20%, the (PS)2-v2 server was significantly better than
(PS)2-original server (p-value is 4.0E-7) and (PS)2-CASP8
(p-value is 6.6E-4) using the paired Student's t-test (Table 4). For each target in CASP8, Table S2 (in Additional file 4) shows the GDT_TS score improvement with contribut-ing components (i.e. multiple templates, multiple mod-els, and template search method) between the (PS)2-v2
and our previous servers.
These 154 TBM targets were also used to evaluate the auto-matic servers participating in CASP8. For the templates selection, the accuracy of identifying the best template of the target protein was used to evaluate the performance of these servers (Figure S2 in Additional file 5). The accura-cies of the (PS)2-v2 server were 54.1% and 75.0% for
iden-tifying the Top 1 templates and Top 10 templates, respectively. In addition, (PS)2-v2 was the rank 6th among
these 72 severs based on GDT_TS scores (Table 5). This
Table 5: Comparing (PS)2-v2 with 71 automatic servers on 154 targets in CASP8
Rank Servers Sum of GDT_TS score
1 Zhang-Server 10870.7 2 RAPTOR 10584.5 3 pro-sp3-TASSER, Phyre_de_novo 10469.3 ~ 10452.9 5 BAKER-ROBETTA, (PS)2-v2, MULTICOM-CLUSTER 10358.9, 10331.4, 10325.8 8 METATASSER 10296.7 ... ... ... ... ... ... 72 mahmood-torda-server 1355.2
Table 4: Comparison the (PS)2-v2 server with (PS)2-original and (PS)2-CASP8 servers on the 154 TBM targets in CASP8 based on
GDT_TS scores
Servers SIa 30% (nb = 40) 20% SI < 30% (n = 47) SI < 20% (n = 67)
Average p-value Average p-value Average p-value
(PS)2-original 82.6 0.0984 67.7 0.0029 44.9 4.0E-7
(PS)2-CASP8 84.3 0.323 70.6 0.0766 51.0 6.6E-4
(PS)2-v2 84.3 - 71.1 - 54.0
-a Sequence identity (SI) between the target and the best template. bn is the number of targets.
server is often able to yield reliable predicted structures (i.e. GDT_TS score = 60%) if the E-value is less than 10-2
(Figure S3 in Additional file 6).
The top-rank five serves (Zhang-Server, RAPTOR, pro-sp3-TASSER, Phyre_de_novo and BAKER-ROBETTA) are better than (PS)2-v2 on 40 hard targets (i.e., LGA_S score < 70%)
(Table S3 in Additional file 7). These serves were much slower than (PS)2-v2 because they often utilized ab initio
methods to build the unaligned loop regions and to gen-erate the models, such as the Poing folding system for Phyre_de_novo server, the chunk-TASSER [38] for pro-sp3-TASSER server, and the Rosetta fragment-assembly methodology [39] for BAKER-ROBETTA server. In the
near future, our (PS)2-v2 server will incorporate ab initio
methods to model long-length loops and hard targets.
Multiple templates for multiple domains
We used the target T0504 as an example to describe (PS)2
-v2 for selecting multiple templates to model protein struc-tures (Figure 5). The (PS)2-v2 server first selected the
53BP1 tandem tudor domains (PDB code 2g3r) as the best template. The template 2g3rA aligned a part of regions (138 residues, residues 10-147) to the target, and the model yielded the GDT_TS scores of 74.2 and 32.2 for the target T0504-D1 and T0504-D2. Since the number of the unaligned residues is 61 (residue 148-208), the (PS)2
-v2 server used unaligned residues to search the new tem-plate for modeling this segment. After search temtem-plate library, (PS)2-v2 selected the PHD finger protein 20-like 1
(PDB code 2eqm) as the template for modeling this unmodeling residues (T0504-D3). The GDT_TS score of this model is 80.7 for the target T0504-D3. The total GDT_TS score improvement is 136.42 when (PS)2-v2
uti-lizes a multiple-template strategy. Conversely, the GDT-TS scores of the (PS)2-original server, using PDB code 2g3r as
the template, are 17.3 (T0504-D1), 48.9 (T0504-D2) and 56.1 (T0504-D3), respectively. For the (PS)2-CASP8
server, the GDT-TS scores using PDB code 2ns2 as the tem-plate are 44.4 (T0504-D1), 25.6 (T0504-D2) and 41.0 (T0504-D3), respectively.
Multiple models and model selection
Figure 6 shows the improvement in GDT_TS scores of (PS)2-v2 by applying a multiple-model strategy and using
the program ProQ for the final model selection. Among these 154 CASP8 targets, (PS)2-v2 improved GDT_TS
scores for 23 targets; conversely, only 4 targets are lightly worse when (PS)2-v2 used a multiple-model strategy. For
the other 127 targets, (PS)2-v2 obtained the same GDT_TS
scores and the total GDT_TS improvement is 145.3. According to the paired Student's t-test (p-value is 0.0045 shown in Table S4 Additional file 8), (PS)2-v2 applying
the multiple-model strategy significantly improved the GDT_TS scores when the sequence identity between the target and the template is less than 20%.
The target T0471 selected from CASP8 was taken as an example to describe the structure modeling of the (PS)2
-v2 server using multiple-model strategy (Figure 7). When the multiple-model strategy was not considered, (PS)2-v2
selected the 2-dehydro-3-deoxyphosphooctonate aldolase (PDB code 2nwr) as the best template with an E-value of 0.055. GDT_TS score of this model is 32.67. If we consid-ered the top-ranking 5 structures (PDB codes 2nwr, 1pea, 1nv8, 1ufr and 1v2d) as the modeling templates, (PS)2-v2
generated 6 alternative target-template alignments for each template, and obtained 30 alignments for this target. The software MODELLER was then applied to generate 30 Comparison the (PS)2-v2 server with (A) (PS)2-original and
(B) (PS)2-CASP8 servers on the 154 TBM targets in CASP8
Figure 4
Comparison the (PS)2-v2 server with (A) (PS)2 -origi-nal and (B) (PS)2-CASP8 servers on the 154 TBM tar-gets in CASP8. (PS)2-v2 yields 99 and 34 higher GDT_TS
scores than (PS)2-original and (PS)2-CASP8, respectively,
among these 154 targets. These three servers have the simi-lar GDT_TS scores when the sequence identity (SI) between the target and template is more than 30% (blue +). (PS)2-v2
outperforms our previous servers when SI is less than 20% (green ×). ! "# $ % &' () (* + , - . ʁ/ 0 1 2 0 1 ʀ - . 3 / 0 1 - .3 2 0 1 4 5 4 6 4 7 4 8 4 9 4 4 4 5 4 6 4 7 4 8 4 94 4 : ; < = < >? @ A B C A D E F > G HI J K L M N O N M PQ R ST R U VW X YZ[\ ] X W 8 ^ _ ʁ` a b c a b ʀ ^ _d ` a b ^ _ d c a b e f
structures for these 30 target-template alignments. Figure 7 shows the best model with the highest LGscores, assess-ing by the program ProQ, for each template. The model generated by the template 1nv8A was selected as the final model, because it had the best LGscore (2.838) among these 30 models. The GDT_TS score of this final model is 61.65. The (PS)2-v2 server using multiple models is often
able to effectively improve accuracies when the E-value between the target and the template is more than 0.01. The average GDT_TS improvements are 8.53 and 2.23, respectively, when the E-value ≥ 0.01 and E-value ≤ 1e-6.
T0409 in CASP8
The target T0409 selected from CASP8 was taken to describe the structure modeling of the (PS)2-v2 server
(Figure 8). The target is the BIG_1156.2 domain of puta-tive penicillin-binding protein MrcA from Nitrosomonas
europaea ATCC 19718. This server yielded the best
GDT_TS score (77.8) among all participating servers for this target.
For the target T0409, the (PS)2-v2 server selected the
C-ter-minal domain of translation initiation factor 5A protein (PDB code 1bkb) from Pyrobaculum aerophilum as the tem-plate [40]. The C-terminal domain is found to be homol-Comparison the (PS)2-v2 server with (PS)2-original and (PS)2-CASP8 servers on the target T0504 in CASP8
Figure 5
Comparison the (PS)2-v2 server with (PS)2-original and (PS)2-CASP8 servers on the target T0504 in CASP8. The (PS)2-CASP8 server uses human spindlin1 (PDB code 2ns2) as the template, conversely, (PS)2-v2 utilizes a
multiple-tem-plate strategy and selects both 53BP1 tandem tudor domains (PDB code 2g3r) and PHD finger protein 20-like 1 (PDB code 2eqm) as templates. (PS)2-v2 significantly outperforms (PS)2-CASP8 on the T0504-D1 and T0504-D3 domains.
! " # $ # % & ' % ( ) * ) + , -( *) . / 0 1 2 3 ) ( 4 5 ( ( 6 7 8 3 ) 9 + : ; 2 0 2 ) + , -( *) <
(PS)2-v2 results for using single-model and multiple-model strategies on 154 targets in CASP8 based on GDT_TS scores Figure 6
(PS)2-v2 results for using single-model and multiple-model strategies on 154 targets in CASP8 based on GDT_TS scores. (PS)2-v2 improves and decreases the
GDT_TS scores for 23 and 4 targets, respectively, when the multiple-model method is utilized. For the other 127 targets, (PS)2-v2 obtains the same GDT_TS scores. The symbols "+",
"䊐" and "×" represent the performance when the sequence
identity (SI) ≥ 30%, between 30% and 20%, and less than 20%, respectively. ! " ! # $% &'() * + ʁ, - . / - . ʀ * +0 , - . * +0 / - .
(PS)2-v2 models the target T0471 in CASP 8 using multiple models
Figure 7
(PS)2-v2 models the target T0471 in CASP 8 using multiple models. This server models T0471 by selecting top-rank-ing five structures (PDB code 2nwrA, 1peaA, 1nv8A, 1ufrA and 1v2dA) as templates ustop-rank-ing S2A2 matrix and PSSM scortop-rank-ing matrices. For each template, (PS)2-v2 generates 5 structures and (D) the final model (1nv8) is identified by the program ProQ
based on LGscore. ! " # $ % & ' ( # ) $* + " , ! " # $* * & ' ( # - $) . - ! " # $- ) - & ' ( # * $* % /0 ! " # $* - & ' ( # $* . 1 ! " # $ ) 2 & ' ( # % $
-An example of the prediction results of the target T0409 from the (PS)2-v2 server
Figure 8
An example of the prediction results of the target T0409 from the (PS)2-v2 server. The alignment and predicted structure of the BIG_1156.2 domain of putative penicillin-binding protein MrcA from Nitrosomonas europaea ATCC 19718 using the (PS)2-v2 server. (A) The alignment between the query and the selected template, translation initiation factor 5A
pro-tein (PDB code 1bkbA), from Pyrobaculum aerophilum. (B) The superposition, the native structure of T0409 (broad, PDB code 3d0f) and the predicted structure (thin). The green blocks are the regions that the predicted structure matches to the native structure. The yellow and purple blocks indicate the shift errors between predicted structure and native structure, the Cα dis-tances between them are <5 Å and >5 Å, respectively.
! " " # # $ # $ % " " & % " $
ogous to the cold-shock protein CspA of E. coli, which has a well characterized RNA-binding fold. The best template reported in the CASP8 website is the yeast exosome core, Rrp44 (PDB code 2vnvD) [41], which contains four domains (CSD1, CSD2, RNB and S1). The S1 domain has the most similar structure to the target T0409-D1. The S1 domain also has a common OB fold characteristic of RNA-binding protein, with five anti-parallel β strands. Figure 8A shows the target-template alignment and the template shares 17.0% sequence identity with the query sequence. Our server could align the five anti-parallel β strands together. Figure 8B shows the superposition of the predicted structure (thin) and the X-ray structure (broad) of the target T0409.
Conclusion
This study presents an automatic server for protein struc-ture predictions by applying numerous enhancements and modifications to the original technique, thereby improving the reliability and applicability. By integrating the S2A2 and PSSM matrixes, the (PS)2-v2 server
seam-lessly blends the amino acid and structural propensities so that they work cooperatively for the template selection and target-template alignments. In addition, our (PS)2-v2
utilizes multiple templates and multiple models for build-ing models and assessbuild-ing models. Experimental results demonstrate that the (PS)2-v2 server is efficient and
effec-tive for template selections and target-template align-ments in template-based modeling. We believe that this server is useful in protein structure prediction and mode-ling, especially in detecting homologous templates with sequence similarity in the twilight zone.
Availability and requirements
Project home page: http://ps2v2.life.nctu.edu.tw Operating system(s): Platform independent Programming language: C, Perl and PHP
Other requirements: JavaScript-enabled web browser Any restrictions to use by non-academics: None
Competing interests
The authors declare that they have no competing interests.
Authors' contributions
Conceived and designed the experiments: CCC and JMY. Performed the experiments and analyzed the data: CCC and JMY. Contributed reagents/materials/analysis tools and wrote the paper: CCC, JKH and JMY.
Additional material
Additional file 1
Program. The (PS)2-v2 program. Click here for file[http://www.biomedcentral.com/content/supplementary/1471-2105-10-366-S1.gz]
Additional file 2
Figure S1. The S2A2 matrix.Click here for file
[http://www.biomedcentral.com/content/supplementary/1471-2105-10-366-S2.pdf]
Additional file 3
Table S1. The summary of 154 TBM targets in CASP8.
Click here for file
[http://www.biomedcentral.com/content/supplementary/1471-2105-10-366-S3.pdf]
Additional file 4
Table S2. The GDT_TS scores of the (PS)2-original, (PS)2-CASP8 and
(PS)2-v2 servers on 154 TBM targets. Click here for file
[http://www.biomedcentral.com/content/supplementary/1471-2105-10-366-S4.pdf]
Additional file 5
Figure S2. Comparison of the (PS)2-v2 server with top-ranking 45 servers
participating in the CASP8 competition for the template selection on 154 TBM targets. The best templates are directly summarized from the CASP8 website http://predictioncenter.org/casp8/.
Click here for file
[http://www.biomedcentral.com/content/supplementary/1471-2105-10-366-S5.pdf]
Additional file 6
Figure S3. The relation between E-values and GDT_TS scores of (PS)2
-v2 for the targets in CASP8. (PS)2-v2 often yields reliable predicted
struc-tures if the E-value is less than 10-2. Click here for file
[http://www.biomedcentral.com/content/supplementary/1471-2105-10-366-S6.pdf]
Additional file 7
Table S3. Comparison of the (PS)2-v2 server and top five servers in
CASP8.
Click here for file
[http://www.biomedcentral.com/content/supplementary/1471-2105-10-366-S7.pdf]
Additional file 8
Table S4. (PS)2-v2 results for using single-model and multiple-model
strategies on 154 targets in CASP8 based on GDT_TS scores.
Click here for file
[http://www.biomedcentral.com/content/supplementary/1471-2105-10-366-S8.pdf]
Acknowledgements
J.-M. Yang was supported by National Science Council and partial support of the ATU plan by MOE. Authors are grateful to both the hardware and software supports of the Structural Bioinformatics Core Facility at National Chiao Tung University.
References
1. Aloy P, Pichaud M, Russell RB: Protein complexes: structure
pre-diction challenges for the 21(st) century. Curr Opin Struct Biol
2005, 15(1):15-22.
2. Pieper U, Eswar N, Davis FP, Braberg H, Madhusudhan MS, Rossi A, Marti-Renom M, Karchin R, Webb BM, Eramian D, et al.:
MOD-BASE: a database of annotated comparative protein struc-ture models and associated resources. Nucleic Acids Res 2006, 34:D291-D295.
3. Schwede T, Kopp J, Guex N, Peitsch MC: SWISS-MODEL: an
automated protein homology-modeling server. Nucleic Acids Res 2003, 31(13):3381-3385.
4. Zhang Y: I-TASSER server for protein 3D structure
predic-tion. BMC Bioinformatics 2008, 9:40.
5. Chivian D, Kim DE, Malmstrom L, Schonbrun J, Rohl CA, Baker D:
Prediction of CASP6 structures using automated Robetta protocols. Proteins 2005, 61:157-166.
6. Zhou HY, Zhou YQ: Fold recognition by combining sequence
profiles derived from evolution and from depth-dependent structural alignment of fragments. Proteins 2005, 58(2):321-328.
7. McGuffin LJ, Jones DT: Improvement of the GenTHREADER
method for genomic fold recognition. Bioinformatics 2003, 19(7):874-881.
8. Rice DW, Eisenberg D: A 3D-1D substitution matrix for protein
fold recognition that includes predicted secondary structure of the sequence. J Mol Biol 1997, 267(4):1026-1038.
9. Zhou HY, Zhou YQ: Single-body residue-level
knowledge-based energy score combined with sequence-profile and sec-ondary structure information for fold recognition. Proteins
2004, 55(4):1005-1013.
10. Qiu J, Elber R: SSALN: An alignment algorithm using
struc-ture-dependent substitution matrices and gap penalties learned from structurally aligned protein pairs. Proteins 2006, 62(4):881-891.
11. Kelley LA, MacCallum RM, Sternberg MJE: Enhanced genome
annotation using structural profiles in the program 3D-PSSM. J Mol Biol 2000, 299(2):499-520.
12. Kawabata T, Nishikawa K: Protein structure comparison using
the Markov transition model of evolution. Proteins 2000, 41(1):108-122.
13. Liu S, Zhang C, Liang SD, Zhou YQ: Fold recognition by
concur-rent use of solvent accessibility and residue depth. Proteins
2007, 68(3):636-645.
14. Lindahl E, Elofsson A: Identification of related proteins on
fam-ily, superfamily and fold level. J Mol Biol 2000, 295(3):613-625.
15. Domingues FS, Lackner P, Andreeva A, Sippl MJ: Structure-based
evaluation of sequence comparison and fold recognition alignment accuracy. J Mol Biol 2000, 297(4):1003-1013.
16. Chen CC, Hwang JK, Yang JM: (PS)2: protein structure
predic-tion server. Nucleic Acids Res 2006, 34:W152-W157.
17. Chen CC, Yang JM, Hwang JK: (PS)2: protein structure
predic-tion server. Eighth Community Wide Experiment on the Critical Assess-ment of Techniques for Protein Structure Prediction 2008:87-88.
18. Pearson WR: Searching Protein-Sequence Libraries -
Com-parison of the Sensitivity and Selectivity of the Smith-Waterman and Fasta Algorithms. Genomics 1991, 11(3):635-650.
19. Wallner B, Elofsson A: Can correct protein models be
identi-fied? Protein Sci 2003, 12(5):1073-1086.
20. Apweiler R, Bairoch A, Wu CH, Barker WC, Boeckmann B, Ferro S, Gasteiger E, Huang HZ, Lopez R, Magrane M, et al.: UniProt: the
Universal Protein knowledgebase. Nucleic Acids Res 2004, 32:D115-D119.
21. Jones DT: Protein secondary structure prediction based on
position-specific scoring matrices. J Mol Biol 1999, 292(2):195-202.
22. Deshpande N, Addess KJ, Bluhm WF, Merino-Ott JC, Townsend-Merino W, Zhang Q, Knezevich C, Xie L, Chen L, Feng ZK, et al.: The
RCSB Protein Data Bank: a redesigned query system and relational database based on the mmCIF schema. Nucleic Acids Res 2005, 33:D233-D237.
23. Kabsch W, Sander C: Dictionary of protein secondary
struc-ture: pattern recognition of hydrogen-bonded and geometri-cal features. Biopolymers 1983, 22:2577-2637.
24. Sali A, Blundell TL: Comparative Protein Modeling by
Satisfac-tion of Spatial Restraints. J Mol Biol 1993, 234(3):779-815.
25. Henikoff S, Henikoff JG: Amino-Acid Substitution Matrices
from Protein Blocks. Proc Natl Acad Sci USA 1992, 89(22):10915-10919.
26. Yang JM, Tung CH: Protein structure database search and
evo-lutionary classification. Nucleic Acids Res 2006, 34(13):3646-3659.
27. Murzin AG, Brenner SE, Hubbard T, Chothia C: Scop - a Structural
Classification of Proteins Database for the Investigation of Sequences and Structures. J Mol Biol 1995, 247(4):536-540.
28. Marti-Renom MA, Madhusudhan MS, Sali A: Alignment of protein
sequences by their profiles. Protein Sci 2004, 13(4):1071-1087.
29. Smith TF, Waterman MS: Identification of common molecular
subsequences. J Mol Biol 1981, 147:195-197.
30. Chen H, Kihara D: A study of quality measures for protein
threading models. BMC Bioinformatics 2001, 2:5.
31. Kraulis PJ: Molscript - a Program to Produce Both Detailed
and Schematic Plots of Protein Structures. J Appl Crystallogr
1991, 24:946-950.
32. Merritt EA, Murphy MEP: Raster3d Version-2.0 - a Program for
Photorealistic Molecular Graphics. Acta Crystallogr Sect D-Biol Crystallogr 1994, 50:869-873.
33. AstexViewer [http://www.astex-therapeutics.com/AstexViewer/
index.php]
34. Lee J, Mandell EK, Tucey TM, Morris DK, Lundblad V: The Est3
pro-tein associates with yeast telomerase through an OB-fold domain. Nat Struct Mol Biol 2008, 15(9):990-997.
35. Wang F, Podell ER, Zaug AJ, Yang YT, Baciu P, Cech TR, Lei M: The
POT1-TPP1 telomere complex is a telomerase processivity factor. Nature 2007, 445(7127):506-510.
36. Yona G, Levitt M: Within the twilight zone: A sensitive
profile-profile comparison tool based on information theory. J Mol Biol 2002, 315(5):1257-1275.
37. Zemla A: LGA: a method for finding 3D similarities in protein
structures. Nucleic Acids Res 2003, 31(13):3370-3374.
38. Zhou HY, Skolnick J: Ab initio protein structure prediction
using Chunk-TASSER. Biophys J 2007, 93(5):1510-1518.
39. Bonneau R, Strauss CEM, Rohl CA, Chivian D, Bradley P, Malmstrom L, Robertson T, Baker D: De novo prediction of
three-dimen-sional structures for major protein families. J Mol Biol 2002, 322(1):65-78.
40. Peat TS, Newman J, Waldo GS, Berendzen J, Terwilliger TC:
Struc-ture of translation initiation factor 5A from Pyrobaculum aerophilum at 1.75 angstrom resolution. Structure with Folding & Design 1998, 6(9):1207-1214.
41. Lorentzen E, Basquin J, Tomecki R, Dziembowski A, Conti E:
Struc-ture of the active subunit of the yeast exosome core, Rrp44: Diverse modes of substrate recruitment in the RNase II nuclease family. Mol Cell 2008, 29(6):717-728.
42. Laskowski RA, Macarthur MW, Moss DS, Thornton JM: Procheck
-a Progr-am to Check the Stereochemic-al Qu-ality of Protein Structures. J Appl Crystallogr 1993, 26:283-291.
43. Xu J, Li M, Kim D, Xu Y: RAPTOR: Optimal protein threading
by linear programming. J Bioinform Comput Biol 2003, 1(1):95-117.
44. Kim D, Xu D, Guo JT, Ellrott K, Xu Y: PROSPECT II: protein
structure prediction program for genome-scale applications. Protein Eng 2003, 16(9):641-650.
45. Cheng JL, Baldi P: A machine learning information retrieval
approach to protein fold recognition. Bioinformatics 2006, 22(12):1456-1463.
