Advertisement
Browse Subject Areas
?

Click through the PLOS taxonomy to find articles in your field.

For more information about PLOS Subject Areas, click here.

  • Loading metrics

Open Access

Peer-reviewed

Research Article

SP5: Improving Protein Fold Recognition by Using Torsion Angle Profiles and Profile-Based Gap Penalty Model

  • Wei Zhang,

    Affiliations Indiana University School of Informatics and Center for Computational Biology and Bioinformatics, Indiana University School of Medicine, Indiana University-Purdue University Indianapolis, Indianapolis, Indiana, United States of America, Institute of Applied Physics and Computational Mathematics, Beijing, People's Republic of China

    ⨯
  • Song Liu,

    Affiliations Department of Biostatistics, Center of Excellence in Bioinformatics & Life Sciences, University at Buffalo, State University of New York, Buffalo, New York, United States of America, Department of Biostatistics, Roswell Park Cancer Institute, Buffalo, New York, United States of America, Howard Hughes Medical Institute Center for Single Molecule Biophysics, Department of Physiology & Biophysics, University at Buffalo, State University of New York, Buffalo, New York, United States of America

    ⨯
  • Yaoqi Zhou

    yqzhou@iupui.edu

    Affiliation Indiana University School of Informatics and Center for Computational Biology and Bioinformatics, Indiana University School of Medicine, Indiana University-Purdue University Indianapolis, Indianapolis, Indiana, United States of America

    ⨯

Abstract

How to recognize the structural fold of a protein is one of the challenges in protein structure prediction. We have developed a series of single (non-consensus) methods (SPARKS, SP2, SP3, SP4) that are based on weighted matching of two to four sequence and structure-based profiles. There is a robust improvement of the accuracy and sensitivity of fold recognition as the number of matching profiles increases. Here, we introduce a new profile-profile comparison term based on real-value dihedral torsion angles. Together with updated real-value solvent accessibility profile and a new variable gap-penalty model based on fractional power of insertion/deletion profiles, the new method (SP5) leads to a robust improvement over previous SP method. There is a 2% absolute increase (5% relative improvement) in alignment accuracy over SP4 based on two independent benchmarks. Moreover, SP5 makes 7% absolute increase (22% relative improvement) in success rate of recognizing correct structural folds, and 32% relative improvement in model accuracy of models within the same fold in Lindahl benchmark. In addition, modeling accuracy of top-1 ranked models is improved by 12% over SP4 for the difficult targets in CASP 7 test set. These results highlight the importance of harnessing predicted structural properties in challenging remote-homolog recognition. The SP5 server is available at http://sparks.informatics.iupui.edu.

Citation: Zhang W, Liu S, Zhou Y (2008) SP5: Improving Protein Fold Recognition by Using Torsion Angle Profiles and Profile-Based Gap Penalty Model. PLoS ONE 3(6): e2325. https://doi.org/10.1371/journal.pone.0002325

Editor: David Jones, University College London, United Kingdom

Received: February 16, 2008; Accepted: April 28, 2008; Published: June 4, 2008

Copyright: © 2008 Zhang et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Funding: This work was supported by NIH (R01 GM 966049 and R01 GM 068530).

Competing interests: The authors have declared that no competing interests exist.

Introduction

Fold recognition refers to recognizing the structural fold of a protein, given its sequence information. Fold recognition is one of the key bottlenecks for protein structure predictions as the protein data bank now appears to contain the complete (or near complete) set for all possible structural folds of proteins, at least for small domain proteins [1], [2].

Recently completed assessment of automated servers for protein structure prediction (CASP 7) [3] reveals the power of post-treatment of models predicted by individual fold recognition methods through consensus predictions (For example, ROBETTA [4], Pmodeller6 [5], Fams-ace [6]) and/or constrained template-fragment recombination and refinement (For example, Chunk-TASSER [7], I-TASSER [8]). The prediction quality of these methods, however, relies heavily on the accuracy of initial models generated by individual fold recognition methods in the first step. Another observation is that the accuracy of top single servers can rival with most consensus methods. Thus, developing and/or improving individual methods are critically important for further advancement of the accuracy of fold recognition and structure prediction.

We have developed a series of single fold-recognition methods (SPARKS, SP2, SP3, SP4) that are based on weighted matching of multiple profiles that include sequence profiles generated from multiple sequence alignment [9], predicted versus actual secondary structures [10], [11], knowledge-based profile (single-body) score function [10], depth-dependent sequence profiles derived from template structures [11], and predicted versus actual solvent accessible surface area [12]. There is a robust improvement of the accuracy and sensitivity of fold recognition as the number of matching profiles increases [10], [11], and [12]. SPARKS, SP3, and SP4 were ranked among the top performers for automatic servers in recent CASP 6 [13], [14] and 7 [12], [3]. This exemplifies the importance and effectiveness of multiple-dimensional use of the structural information of templates in developing fold-recognition techniques.

In this paper, we introduce the fifth “dimension” for fold recognition by incorporating predicted backbone torsion angles (SP5). The backbone torsion angles (φ and ψ) are two rotation angles about the Cα – N bond (φ) and the Cα – C bond (ψ). Because the polypeptide backbone of a protein is a linked sequence of rigid planar peptide groups, these two angles essentially determine the backbone conformation of proteins. While a three-state classification of secondary structures is a coarse-grained one-dimensional representation of local backbone conformation, backbone torsion angles encode the backbone tertiary structure, at least in principle.

Traditionally, dihedral torsion angles are predicted as a few discrete states based on local (fragment) structural patterns using either machine-learning techniques or classification schemes [15]–[22]. However, there were only a few limited applications of predicted angle states to fold recognition [18] and sequence alignment [23]. The former uses torsion-angle states as a replacement of simple three-state secondary structures to build an iterated alignment hidden Markov model [18]. The latter [23] predicts angle states by hidden Markov model and employs the predicted angles to build structural context-based substitution matrices. Here, we propose to match predicted and actual torsion angles as a new profile term in a multi-dimensional profile-profile alignment. This represents a novel use of predicted torsion angles as a complementary to rather than a replacement of secondary structures for fold recognition. The angel profile used in this work is built on a recent advancement in real-value prediction of torsion angles [24]. By taking advantage of angle periodicity and using integrated neural networks, we have obtained ten-fold-cross-validated mean absolute errors of 38° for ψ and 25° for φ [24]. This accuracy of real-value prediction was found comparable to or more accurate than those based on multi-state classification of the φ – ψ map.

In SP4, the effect of solvation was taken into consideration by matching the predicted and actual solvent accessibility (SA). The SA profiles are based on two states (exposed and buried) classified according to an arbitrary threshold of 25%. The two-state classification increases the accuracy of prediction by reducing number of states in SA. This is at the cost of losing the detailed fluctuation pattern of SA along the sequence. We recently have developed method (called Real-SPINE) for real value SA prediction, which yields a 10-fold cross-validated Pearson's correlation coefficient (PCC) of 0.74 between predicted and actual solvent SA [25]. We thus have updated the original two-state SA profile with the new real-value one in developing SP5 scoring function.

In addition to the torsion angle and real-value SA term, we will introduce a new variable gap-penalty model to replace the original constant gap-penalty model. The new model is based on insertion and deletion probability profiles generated from PSIBLAST. Several studies [26]–[28] have indicated the usefulness of these context-dependent profiles for improving alignment accuracy. Here, we propose an implementation by using insertion and deletion probability profiles to a fractional power.

The above-proposed algorithm leads to the new method called SP5. SP5 is tested in two alignment benchmarks and two structure-modeling benchmarks. Results suggest a significant improvement of SP5 over SP3 and SP4 in fold recognition.

Results

Parameter Optimization by the PREFAB Benchmark

Weight factors and gap parameters in SP3 and SP4 were optimized by using Prosup benchmark [38]. In this study, we use PREFAB 4.0 to optimize SP5 parameters [39]. We use PREFAB because its reference alignment is made from the consensus of two separate structural alignment programs (CE [40] and FSSP [41]) rather than one in Prosup. Ninety one pairs of proteins are randomly selected from PREFAB benchmark, with sequences identity less than 30% from each other. We optimized the parameters for SP5 (with new profile-based gap model) by maximizing the percent of matches between the reference alignment in PREFAB and the alignment made SP5. The optimization is done by sequential grid-search until further iterations do not improve the alignment accuracy [11]. The final parameters used are w0 = 5.6, w1 = 0.68, sshift = −0.27, w2ndary = 0.52, wstruc = 0.46, wsa = 2.3, wΔ = 1.33, with the accuracy of one-to-one match 62.3%.

Testing Alignment Accuracy by ProSup and SALIGN Benchmarks

The alignment accuracy of the methods trained by PREFAB benchmarks is tested by the ProSup and SALIGN benchmarks. Prosup benchmark, prepared by Sippl's group, consists of 127 pairs of proteins with alignment by structural alignment program Prosup [38]. SALIGN benchmark [42] contains 200 selected pairs with an average pair sharing 20% sequence identity and 65% of structurally equivalent Cα atoms superposed with an rmsd of 3.5 Å [42]. Reference alignment is obtained from the structural alignment obtained from the TMalign program [43] [i.e., TM overlap]. The sequence identity between PREFAB training set and test sets SALIGN and Prosup are 18% and 20%, respectively.

Table 1 shows the alignment accuracy of different methods given by different benchmarks along with the standard errors estimated by bootstrap simulation on 10,000 re-sampling of the data. There is a consistent improvement from SP3, SP4 to SP5. The absolute changes range from 1.9% to 2.4% (3.4%) from SP4 (SP3) to SP5 while the relative increases are between 3–5% (5–6%) [SP5 relative to SP4 (SP3)]. These changes are significantly greater than the estimated standard errors. The improvement is remarkable considering the fact that ProSup benchmark was used as the training set to optimize the parameters of SP3 [11] and SP4 [12].

thumbnail
Download:
Table 1. The alignment accuracies for Prosup and SALIGN benchmark.

https://doi.org/10.1371/journal.pone.0002325.t001

Testing Fold Recognition with Lindahl Benchmark

The ultimate purpose of improving alignment is to make more accurate fold recognition and structure prediction. Lindahl Benchmark is a large data set of 976 proteins, with 555, 434, and 321 pairs of proteins in the same family, superfamily, and fold, respectively [44]. However, DSSP [30] failed to produce results for 9 proteins. Thus, the actually used proteins in this study are 967 and the number of proteins in family, superfamily and fold is 550, 430, and 317, respectively. Here, the fold recognition sensitivity of each method is tested by aligning each protein with the rest 966 proteins, and checking whether or not the method can recognize the member of same family, superfamily or fold as the first rank or within the top 5 ranks. Thus, the benchmark tests both modeling accuracy and ranking methods of proposed methods.

Table 2 shows the fraction of correctly recognized match of proteins in the same family, superfamily, fold as first rank or within top 5 rank of the templates given by various SP methods and several other methods. Although many published methods have been applied to this benchmark [45]–[47], [10], we only list most recent ones [10], [11], [48], [12]. This is because of the time dependent nature of sequence database for sequence profiles. For facilitating the comparison within SP methods, we used original sequence profiles from Ref. [11].

thumbnail
Download:
Table 2. The success rate for recognizing proteins within the same family, superfamily, or fold in the Lindahl benchmark.

https://doi.org/10.1371/journal.pone.0002325.t002

Table 2 indicates that the improvement over SP3 and SP4 in success rate of fold recognition by SP5 exists in all three levels (family, superfamily, and fold). The largest improvement over SP4 is observed in fold level (7% absolute increase in Top 1 and 5% absolute increase for the best in Top 5; 22% relative increase in Top 1, 9.5% in top 5). This is somewhat expected because the method was trained for remote homolog recognition (structurally similar protein with less than 30% sequence identity, PREFAB benchmark). Again the relative improvement of SP5 over SP3 and SP4 is significantly larger than the standard errors estimated from bootstrap simulations. We further removed 43 proteins that have >30% sequence identity with the training sequences in the PREFAB benchmark. Their effect on the final result is negligible. For comparison, we also include the results of PSIBLAST [9], SPARKS [10], HHsearch / HHpred [27] and FOLDpro [48]. The performance of SPAKRS and Foldpro was from Ref. [10] and Ref. [48], respectively. We further performed PSIBLAST and HHpred locally with their default parameters. Among all methods listed [9]–[12], [27], [48], SP5 method has the highest success rate on the fold level (both first and top 5 ranks) and the superfamily for the first rank.

Above success rates of matching sequences within the same SCOP classification are based on somewhat subjective SCOP definition of family, superfamily and fold [49]. A more direct measurement of accuracy is to calculate the accuracy of the first-ranked model built from the fold-recognition alignment. The model is first built by transferring the Cα coordinates of the template structures to the aligned residues in the query sequence. The constructed model is then assessed by using the MaxSub score between the model and the known native structure. MaxSub score [37] between the predicted (model) structure and the native structure is a measure of similarity between 0.0 (no similarity) and 1.0 (perfect similarity). The value is calculated by searching the largest subset of well-superimposed residues (≤3.5 Å). Table 3 reports the MaxSub scores for the models built by SP3, SP4 and SP5 methods averaged over the number of proteins. Again SP5 improves over SP4 and SP3 in all levels. The relative improvement of SP5 over SP4 in MaxSub score is 1.4%, 3.1% and 32.2% in family, superfamily and fold levels, respectively.

thumbnail
Download:
Table 3. The model quality of top-1 ranked models in Lindahl benchmark per protein.

https://doi.org/10.1371/journal.pone.0002325.t003

CASP7 test set

We use CASP 7 targets [50] as an additional test set for SP5 method. The test set consists of 95 targets and was released between May and July of 2006. The 95 targets were officially classified into 109 template-based-modeling (TBM) domains and 19 free-modeling (FM) domains, based on whether or not the structurally similar template (deposited in PDB) had been identified and used in prediction.

We test SP3, SP4 and SP5 methods on the CASP7 test set. The template library for SP methods was built in the same way. This was done by using the 40% representative domains of SCOP 1.61. The entire chains of multiple-domain proteins are also contained in the library. The library was then updated with new proteins released after SCOP 1.61 if they have less than 40% sequence identity with the sequences already in the library. To make a strict test, we only include template proteins released before May of 2006 for this test, and we also excluded the templates with sequence identity >20% to the query. The performance of different SP method is evaluated by the Maxsub score of the first ranked Cα model, which is transferred from the alignment.

Table 4 compares the model quality predicted by SP3, SP4, and SP5. Overall, there is a consistent 3% (5% to 6%) improvement from SP5 to SP4 (SP3) for the CASP 7 targets regardless the evaluation based on domains or full chains. For the 109 TBM domains, SP5 is 3% (6%) better than SP4 (SP3). For the most difficult free-modeling targets, there is a 12% improvement from SP4 to SP5. This pattern of improvement is consistent with that from Lindahl benchmark. That is, the most significant improvement from SP4 to SP5 is on the most challenging targets.

thumbnail
Download:
Table 4. The model quality of top-1 ranked models for CASP7 test set.

https://doi.org/10.1371/journal.pone.0002325.t004

Discussion

This paper reports several significant changes over previously developed SP method: the torsion-angle term for profile-profile matching, real-value-based SA profile, and variable gap-penalty model based on fractional-powered insertion/deletion profiles. We showed that by integrating these new features with existing sequence-derived profile, secondary structure profile, residue depth-dependent structure-based profile, the new method SP5 makes a robust improvement over previously developed SP serial methods. Comparing with SP3 and SP4, there is a 2–6% absolute improvement in one-to-one match of alignment accuracy depending on benchmarks. Application of SP5 to the large Lindahl benchmark reveals 1%, 2% and 7% improvements over SP4 in success rates in recognizing proteins within the same family, superfamily and fold, respectively. The improvement in recognition leads to 1%, 3% and 32% improvement in modeling accuracy based on the top-1 ranked, family, superfamily and fold-level models, respectively. Additional test on CASP 7 targets yields 3–6% improvement in 109 template-based modeling targets and 12% improvement in 19 free-modeling targets. Thus, SP5 marks a significant improvement over SP3 and SP4 in fold-recognition, as designed.

This paper represents a full exploitation of predicted torsion-angles for fold recognition. Previous similar studies [18], [23] are limited to view discrete torsion-angle states as an expansion of secondary structures. This paper, however, treats predicted angles as complementary information to predicted three-state secondary structures. The two quantities are complementary because three-state secondary structures represent a coarse-grained description of local structures while torsion angles contain detailed local and nonlocal structural information if they are predicted accurately. Indeed, our limited initial test indicates that removing secondary structures from SP5 will reduce its alignment accuracy. Obviously, the success of SP5 is made possible because of reasonably accurate real-value prediction of torsion angles [24].

Recent progress in sequence alignment and structure prediction has suggested the importance of variable gap penalties in protein sequence alignment [51]. Different form of context (either structure or sequence context or both)-dependent gap-penalty model has been proposed [52], [53]. Employing fractional-powered gap insertion/deletion profiles is another new feature introduced in SP5. While these insertion/deletion profiles were used, previously [26]–[28], our trial-and-error analysis indicates that the fractional-powered gap insertion/deletion profiles with a power of 0.1 seem to be more suitable for improving alignment accuracy. However, more systematic comparative studies are needed to check if any other functional forms are more appropriate.

To analyze the usefulness of the new gap model, we made a version of SP5 with the previously used gap model and found that new gap model leads to a small but positive increase in alignment accuracy (0.5% in PREFAB, 1.5% in ProSup and 0.1% in SALIGN). Thus, the main contribution for improved ability in fold recognition by SP5 is due to introduction of torsion angles.

SP3 and SP4 were among the top performers in automatic servers in CASP 6 and 7 [13], [12]. It is noted that in CASP7, SP3 scored higher than SP4 according to GDT-HA, TMscore, and AL0 for all targets. A close examination [12] indicates that SP4 is slightly more accurate than SP3 in hard targets (FM category), but slightly worse than SP3 in other targets (TBM category). This is perhaps because all parameters were optimized for fold recognition targets. On the other hand, SP4 performs consistently better than SP3 at both FM and TBM categories if the cumulative Z-score is used [12]. The development of SP5 continues our emphasis on searching a more sensitive method for fold recognition. Significant improvement of SP5 over SP4 and SP3 indicates that SP5 is among the most accurate automatic servers for fold recognition.

In the SP serial methods, the alignment generated for fold recognition is used directly in modeling. It is quite possible that a separate alignment method optimized for modeling may further improve the accuracy of predicted model. This will be a subject of future studies.

Methods

Alignment Score

The alignment score of SP5 for aligning query position i with the template position j is(1)with four weight parameters (wstruc, w2ndary, wsa, and wΔ) and a constant shift sshift. This score represents weighted matching of five profiles that are described in detail below.

The first term in Eq. (1) is the profile-profile comparison between the sequence profile from the query sequence and that from the template sequence. is the sequence-derived frequency profile of the query sequence, and are the sequence-derived log odd profile of the template sequence and that of query sequence, respectively. These sequence profiles are constructed by three iterations of PSIBLAST [9] searching (E value cutoff 0.001) against non-redundant (NR) sequence database, which was filtered to remove low-complexity regions, transmembrane regions, and coiled-coil segments [29].

The second term in Eq. (1) compares the sequence profile from the query sequence and that derived from the template sequence (sequence profiles that would “fit” to the structure). is a depth-dependent sequence profile generated from the sequences of those structural fragments that are similar to 9-residue segment structures of the template [11].

The third term in Eq. (1) measures the difference between the predicted secondary structure of the query sequence and the actual secondary structure of the template. δsi,sj is a simple function of the secondary structure element si of the query at sequence position i and sj of the template at sequence position j. δsi,sj = 1 if si = sj and δsi,sj = −1 if si≠sj. We use a three-state definition of secondary structures (H for helix, E for strand, and C for coil). The secondary structures of templates are from DSSP [30].We have used the convention: (H, G, I) →H, (E, B) →E, and others →C. The secondary structure for query sequences is predicted by SPINE [31]. The first three terms constitute the method SP3 [11] except that PSIPRED [29] rather than SPINE [31] was used in SP3 to predict the secondary structure of the query sequence. DSSP [30] is used for analyzing template structures because SPINE was trained based on the DSSP definition of secondary structures.

The fourth term in Eq. (1) is the matching score between the predicted solvent accessibility of the query sequence and solvent accessibility of the template structure. sa(i) and sa(j) are the predicted residue solvent accessibility of query sequence and that of the template structure, respectively. The residue solvent accessibilities of query sequence are predicted by Real-SPINE [25] while residue solvent accessibilities of template structures are calculated from DSSP [30] and normalized by unfolded solvent accessible surface areas [32]. The first four terms constitute the method SP4 [33] except that in SP4, PSIPRED [29] rather than SPINE [31] was employed to predict the secondary structure of the query sequence, and the real values of solvent accessibility from Real-SPINE [25] rather than two-state classifications by SABLE [34] are used to predict the residue solvent accessibility of the query sequence.

The fifth term in Eq. (1) is a new addition in SP5. It characterizes the difference between predicted angles (ψ(i) and φ(i)) of the query sequence and actual angles (ψ(j) and φ(j)) of the template structure withReal values of angles for the query sequence are from Real-SPINE 2.0 [24] while these angles are calculated by DSSP [30] for the template structure. Real-SPINE 2.0 is a method for real-value prediction of torsion angles by using back-propagation neural networks trained with a sliding 21-residue window of sequence profiles, representative amino acid properties, and predicted secondary structures. The ten-fold-cross-validated mean absolute errors are 38° for ψ and 25° for φ, respectively.

Profile-based Gap Model

SP3 [11] and SP4 [12] employ a simple secondary-structure dependent gap penalty. No gaps are allowed if si = sj = α (helix) or si = sj = β sheet). The gap opening (w0) and gap extension (w1) penalties are applied to other regions. In this paper, we construct a profile-based gap model from the multiple sequence alignment made by PSIBLAST [9]. The multiple sequence alignment allows us to calculate the probability of deletion at sequence position i, , and the probability of insertion at sequence position i, , and where , , and N are number of deletions in sequence position i, number of insertions in sequence position i, and total number of aligned sequences, respectively.

Thus, we have four profiles: two for query sequences and two for template sequences (, , , and )

The gap penalty is calculated as follows. We still use w0 as the gap opening penalty. The extension gap penalty is modified by for residue i in the query sequence that is aligned with a gap after residue j in template. Similarly, the extension gap penalty is modified by or residue j in template that is aligned with a gap after residue i in query. Here, w1 is a to-be-optimized weight factor. Usually, is an energetic term. Here, we use rather than to avoid singularity at . We set γ = 0.1 by trials and errors.

Dynamic Programming and Template Ranking

Similar to SP3 and SP4, we used the Smith-Waterman local alignment algorithm [35] to optimize the score that matches the query profiles with template profiles based on Eq. (1) with the revised gaping method described above. Note that the optimization of alignment is to minimize the total alignment score due to the negative signs in Eq. (1).

The templates are ranked based on the difference score between the raw alignment score and the reverse alignment raw score in which the alignment is made with the reversed query sequence [36]. The results of fold-recognition alignment are used to build Cα models based on native template structure. This is done by directly transferring the Cα coordinates of the template structures to the aligned residues in the query sequence. If there is no structural similarity between first two models (defined as zero MaxSub score [37]), templates will be re-ranked by the greater one of two Z-scores, which are calculated based on the raw alignment score normalized by the full alignment length and the non-end-gap alignment length, respectively. Here, the Z-score for a template i is given by , where ave and sd denotes the average and standard deviation of normalized score for all the templates. This ranking mechanism was based on an empirical observation. We found that ranking based on the difference score between the raw alignment score and the reverse alignment raw score works well only if there is some structural similarity between the top-two ranked models (i.e. a significant structural cluster detected). Otherwise, ranking based on Z-scores works better [11].

Author Contributions

Conceived and designed the experiments: YZ. Performed the experiments: WZ SL. Analyzed the data: WZ SL. Wrote the paper: YZ WZ SL.

References

  1. 1. Kihara D, Skolnick J (2003) The PDB is a covering set of small protein structures. J Mol Biol 334: 793–802.
  2. 2. Zhang Y, Hubner IA, Arakaki AK, Shakhnovich E, Skolnick J (2006) On the origin and completeness of single domain structures. Proc Natl Acad Sci 103: 2605–2610.
  3. 3. Battey JND, Kopp J, Bordoli L, Read RJ, Clarke ND, et al. (2007) Automated server predictions in CASP7. Proteins 69: (Suppl 8)68–82.
  4. 4. Chivian D, Kim DE, Malmstrom L, Bradley P, Robertson T, et al. (2003) Automated prediction of CASP-5 structures using the robetta server. Proteins 53: (Suppl 6)524–533.
  5. 5. Wallner B, Larsson P, Elofsson A (2007) Pcons.net: protein structure prediction meta server. Nucleic Acids Res 35: W369–W374.
  6. 6. Terashi G, Takeda-Shitaka M, Kanou K, Iwadate M, Takaya D, et al. (2007) Fams-ace: A combined method to select the best model after remodeling all server models. Proteins 69: (Suppl 8)98–107.
  7. 7. Zhou H, Pandit SB, Lee SY, Borreguero J, Chen H, et al. (2007) Analysis of TASSER-based CASP7 protein structure prediction results. Proteins 69: (Suppl 8)90–97.
  8. 8. Zhang Y (2007) Template-based modeling and free modeling by I-TASSER in CASP7. Proteins 69: (Suppl 8)108–117.
  9. 9. Altschul SF, Madden TL, Schaffer AA, Zhang J, Zhang Z, et al. (1997) Gapped BLAST and PSI-BLAST: a new generation of protein database search programs. Nucl Aci Res 25: 3389–3402.
  10. 10. Zhou H, Zhou Y (2004) Single-body residue-level knowledge-based energy score combined with sequence-profile and secondary structure information for fold recognition. Proteins 55: 1005–1013.
  11. 11. Zhou H, Zhou Y (2005) Fold recognition by combining sequence profiles derived from evolution and from depth-dependent structural alignment of fragments. Proteins 58: 321–328.
  12. 12. Liu S, Zhang C, Liang S, Zhou Y (2007) Fold recognition by concurrent use of solvent accessibility and residue depth. Proteins 68: 636–645.
  13. 13. Zhou H, Zhou Y (2005) SPARKS 2 and SP3 servers in CASP 6. Proteins 61: (Suppl. 7)152–156.
  14. 14. Tress M, Ezkurdia I, Grana O, Lopez G, Valencia A (2005) Assessment of predictions submitted for the CASP6 comparative modeling category. Proteins 61: (Suppl 7)27–45.
  15. 15. Kang HS, Kurochkina NA, Lee B (1993) Estimation and use of protein backbone angle probabilities. J Mol Biol 229: 448–460.
  16. 16. Bystroff C, Thorsson V, Baker D (2000) HMMSTR: a hidden Markov model for local sequence-structure correlations in proteins. J Mol Biol 301: 173–190.
  17. 17. deBrevern AG, Etchebest C, Hazout S (2000) Bayesian probabilistic approach for predicting backbone structures in terms of protein blocks. Proteins 41: 271–287.
  18. 18. Karchin R, Cline M, Mandel-Gutfreund Y, Karplus K (2003) Hidden Markov models that use predicted local structure for fold recognition: alphabets of backbone geometry. Proteins 51: 504–514.
  19. 19. de Brevern AG, Benros C, Gautier R, Valadie H, Hazout S, et al. (2004) Local backbone structure prediction of proteins. In Silico Biol 4: 31.
  20. 20. Kuang R, Lesliei CS, Yang A.-S (2004) Protein backbone angle prediction with machine learning approaches. Bioinformatics 20: 1612–1621.
  21. 21. Mooney C, Vullo A, Pollastri G (2006) Protein structural motif prediction in multidimensional phi-psi space leads to improved secondary structure prediction. J Comput Biol 13: 1489–1502.
  22. 22. Zimmermann O, Hansmann UHE (2006) Support vector machines for prediction of dihedral angle regions. Bioinformatics 22: 3009–3015.
  23. 23. Huang YM, Bystroff C (2006) Improved pairwise alignments of proteins in the twilight zone using local structure predictions. Bioinformatics 22: 413–422.
  24. 24. Xue B, Dor O, Faraggi E, Zhou Y (2008) Real-value prediction of backbone torsion angles. Proteins 70: 10.1002/prot.21940.
  25. 25. Dor O, Zhou Y (2007) Real-SPINE: An integrated system of neural networks for real-value prediction of protein structural properties. Proteins 68: 76–81.
  26. 26. Lesk AM, Levitt M, Chothia C (1986) Alignment of the amino acid sequences of distantly related proteins using variable gap penalties. Protein Engineering 1: 77–78.
  27. 27. Soding J, Biegert A, Lupas AN (2005) The HHpred interactive server for protein homology detection and structure prediction. Nucleic Acids Res 33: W244–W248.
  28. 28. Markstein P, Xu Y, editors. (2007) Improvement in protein sequence-structure alignment using insertion/deletion frequency arrays Vol. 6,. San Diego: University of California. World Scientific.
    • 29. Jones DT (1999) Protein secondary structure prediction based on position-specific scoring matrices. J Mol Biol 292: 195–202.
    • 30. Kabsch W, Sander C (1983) Dictionary of protein secondary structure: Pattern recognition of hydrogen-bonded and geometrical features. Biopolymers 22: 2577–2637.
    • 31. Dor O, Zhou Y (2007) Achieving 80% ten-fold cross-validated accuracy for secondary structure prediction by large-scale training. Proteins 66: 838–845.
    • 32. Chothia C (1976) The nature of the accessible and buried surface in proteins. J Mol Biol 105: 1–14.
    • 33. Zhou H, Xue B, Zhou Y (2007) DDOMAIN: Dividing structures into domains using a normalized domain-domain interaction profile. Protein Sci 16: 947–955.
    • 34. Adamczak R, Porollo A, Meller J (2005) Combining prediction of secondary structure and solvent accessibility in proteins. Proteins 59: 467–475.
    • 35. Smith TF, Waterman MS (1981) Identification of common molecular subsequences. J Mol Biol 147: 195–197.
    • 36. Karplus K, Barrett C, Hughey R (1998) Hidden Markov models for detecting remote protein homologies. Bioinformatics 14: 846–856.
    • 37. Siew N, Elofsson A, Rychlewski L, Fischer D (2000) Maxsub: an automated measure for the assessment of protein structure prediction quality. Bioinformatics 16: 776–785.
    • 38. Domingues FS, Lackner P, Andreeva A, Sippl MJ (2000) Structure-based evaluation of sequence comparison and fold recognition alignment accuracy. J Mol Biol 297: 1003–1013.
    • 39. Edgar RC (1994) MUSCLE: multiple sequence alignment with high accuracy and high throughput. Nucl Acids Res 32: 1792–1797.
    • 40. Shindyalov IN, Bourne P (1998) Protein structure alignment by incremental combinatorial extension (CE) of the optimal path. Protein Eng 11: 739–747.
    • 41. Holm L, Sander C (1994) The FSSP database of structurally aligned protein fold families. Nucl Acids Res 22: 3600–3609.
    • 42. Marti-Renom MA, Madhusudhan M, Sali A (2004) Alignment of protein sequences by their profiles. Protein Sci 13: 1071–1087.
    • 43. Zhang Y, Skolnick J (2005) TM-align: a protein structure alignment algorithm based on the TM-score. Nucleic Acids Res 33: 2302–2309.
    • 44. Lindahl E, Elofsson A (2000) Identification of related proteins on family, superfamily and fold level. J Mol Biol 295: 613–625.
    • 45. Shi J, Blundell TL, Mizuguchi K (2001) FUGUE: sequence-structure homology recognition using environment-specific substitution tables and structure-dependent gap penalties. J Mol Biol 310: 243–257.
    • 46. Xu J, Li M, Lin G, Kim D, Xu Y (2003) Protein structure prediction by linear programming. Pacific Symp Biocomputing 8: 264–275.
    • 47. Kim D, Xu D, Guo J, Ellrott K, Xu Y (2003) PROSPECT II: Protein structure prediction program for the genome-scale. Protein Eng 16: 641–650.
    • 48. Cheng J, Baldi P (2006) A machine learning information retrieval approach to protein fold recognition. Bioinformatics 22: 1456–1463.
    • 49. Murzin AG, Brenner SE, Hubbard T, Chothia C (1995) SCOP: a structural classification of proteins database for the investigation of sequences and structures. J Mol Biol 247: 536–540.
    • 50. Moult J, Fidelis K, Kryshtafovych A, Rost B, Hubbard T, et al. (2007) Critical assessment of methods of protein structure prediction - Round VII. Proteins 69: (Suppl 8)3–9.
    • 51. Dunbrack RL Jr (2006) Sequence comparison and protein structure prediction. Curr Opin Struct Biol 16: 374–384.
    • 52. Goonesekere N, Lee B (2004) Frequency of gaps observed in a structurally aligned protein pair database suggests a simple gap penalty function. Nucl Acids Res 32: 2838–2843.
    • 53. Madhusudhan M, Marti-Renom MA, Sali A (2006) Variable gap penalty for protein sequence-structure alignment. Protein Eng Des Sel 19: 129–133.
    • 54. Needleman SB, Wunsch CD (1970) A general method applicable to the search for similarities in the amino acid sequence of two proteins. Journal of molecular biology 48: 443–453.
    • 55. Fischer D, Elofsson A, Rice DW, Eisenberg D (1996) Assessing the performance of inverted protein folding methods by means of an extensive benchmark. Proceeding of the First Pacific Symposium on Biocomputing. pp. 300–318.