Incorporating Ab Initio energy into threading approaches for protein structure prediction
- Mingfu Shao^{1},
- Sheng Wang^{2},
- Chao Wang^{1},
- Xiongying Yuan^{1},
- Shuai Cheng Li^{3},
- Weimou Zheng^{2}Email author and
- Dongbo Bu^{1}Email author
https://doi.org/10.1186/1471-2105-12-S1-S54
© Shao et al; licensee BioMed Central Ltd. 2011
Published: 15 February 2011
Abstract
Background
Native structures of proteins are formed essentially due to the combining effects of local and distant (in the sense of sequence) interactions among residues. These interaction information are, explicitly or implicitly, encoded into the scoring function in protein structure prediction approaches—threading approaches usually measure an alignment in the sense that how well a sequence adopts an existing structure; while the energy functions in Ab Initio methods are designed to measure how likely a conformation is near-native. Encouraging progress has been observed in structure refinement where knowledge-based or physics-based potentials are designed to capture distant interactions. Thus, it is interesting to investigate whether distant interaction information captured by the Ab Initio energy function can be used to improve threading, especially for the weakly/distant homologous templates.
Results
In this paper, we investigate the possibility to improve alignment-generating through incorporating distant interaction information into the alignment scoring function in a nontrivial approach. Specifically, the distant interaction information is introduced through employing an Ab Initio energy function to evaluate the “partial” decoy built from an alignment. Subsequently, a local search algorithm is utilized to optimize the scoring function.
Experimental results demonstrate that with distant interaction items, the quality of generated alignments are improved on 68 out of 127 query-template pairs in Prosup benchmark. In addition, compared with state-to-art threading methods, our method performs better on alignment accuracy comparison.
Conclusions
Incorporating Ab Initio energy functions into threading can greatly improve alignment accuracy.
Keywords
Introduction
Protein structure determination is critical for understanding protein functions, and also highly relevant with therapeutics and drugs design. Computational prediction methods for protein structure play important roles due to the speed of experimental determination methods cannot catch up with that of generation of protein primary sequences by genome projects. Computational protein structure prediction methods can be categorized into free modeling (FM) and template-based modeling (TBM). Specifically, for the protein without structural analogs in the template database, the structural conformation has to be built from the scratch; while for the proteins having structural analogs, the key step is to identify an accurate alignment between the query sequence and a template with known structure.
Both Ab Initio and threading approaches employ scoring functions to capture interactions among residues in an explicit or implicit manner. In essence, protein folding is the combining effects of local interactions and distant interactions among residues. Specifically, local interactions lead to local structural motifs, while non-local interactions arrange local structural motif to form native-like structures.
The Ab Initio approaches for free modeling attempt to find a structural conformation with the lowest energy. Typically, local interactions are described via short structural fragments while nonlocal interactions are captured via an energy function. Various energy functions [1–5] have been proposed, and can be categorized into two classes, i.e., knowledge-based and physics-based. Compared with physics-based energy functions, knowledge-based energy functions are more attractive since they are easy to use and understand. In addition, distance-dependent potentials perform better than distance-independent ones [6].
- 1.
The scoring function does not contain any non-local interaction information explicitly. For example, FASTA [7], BLAST [8], and PSI-BLAST [9] assume independence among residues at different positions while HMMer [10] and HHpred [11] apply Hidden Markov Model to introduce the transition information between adjacent residues into scoring function. Since only local information is taken into consideration in their scoring functions, dynamic programming is a natural technique to obtain a global optimal solution.
- 2.
The scoring function captures non-local interaction information via contact preference. That is, if a pair of residues in the query sequence are aligned to the two ends of an interaction, then this pair will be given a score according to a contact preference matrix. PROSPECT [12] and RAPTOR [13] implemented this kind of energy function and demonstrated the improvements of prediction accuracy. However, the following features of non-local interactions were not taken into consideration explicitly: (i) it is more accurate to describe pairwise interactions in distance-dependent manner than distance-independent ways; and (ii) besides distance, the orientation angles involved in dipole–dipole interactions have also been proved to be useful to discriminate native structures.
The purposes of the study is to investigate whether threading results can be improved through incorporating Ab Initio energy function. Distant interactions are usually described in a more accurate manner in Ab Initio energy function. For example, dDFIRE [6] employs distance-dependent pair-wise interaction rather than distance-independent one. Encouraging progress has been observed in structure refinement where Ab Initio energy function is employed to refine template structure to be more native-like. It is interesting whether Ab Initio energy function improves alignment generating.
In addition, when the global structural information is incorporated, effective algorithms such as dynamic programming do not work any more: if all pairwise interactions are added into scoring function, the optimization problem becomes NP-hard [14]. A variety of techniques, such as integer linear programming [13] and divide and conquer [15] have been proposed to solve this problem. In this study, we propose an efficient, local search based method to identify optimal alignments. Comparing with existing methods [13, 15], which are designed specifically for scoring functions consisting of distant-independent pairwise interaction alone as their global item, our method is more general and can be used to optimize any kind of scoring functions.
Scoring model
The scoring function to assess an alignment A consists of local item L(A) and distant item G(A), i.e., score(A) = ω_{ L }L(A) + G(A), where ω_{ L } denotes weight of local item.
Local item is the weighted sum of mutation score S_{ m }, secondary structure compatibility score S_{ ss }, solvent accessibility score S_{ sa }, gap penalty score S_{ g }, and structural segment compatibility score S_{ CLE }[16], i.e., L(A) = ω_{ m }S_{ m }(A) + ω_{ ss }S_{ ss }(A) + ω_{ CLE }S_{ CLE }(A) + ω_{ sa }S_{ sa }(A) + ω_{ g }S_{ g }(A), S_{ g }(A) = ω_{ go }GO + ω_{ ge }GE, where GO and GE are the number of gap open and gap extending, respectively. The weight of these items are to be determined via training on SALIGN benchmark.
The global item G(A), which contains the nonlocal interaction information implicitly, is captured by the dDFIRE energy over a “partial” decoy corresponding to the alignment A. An ideal way to measure non-local interaction is to calculate dDFIRE energy over a full-length decoy. However, it is usually time-consuming to obtain full-length decoy through running structure-generating tools such as MODELLER [17]. Thus, this strategy is unacceptable since we usually need to sample thousands of alignments. Here, we employ an alternative method to build a partial “decoy” from the alignment. Specifically, only the aligned residues are kept with their coordinates simply copied from the corresponding residues in the template.
This section are organized as follows: We first verify that dDFIRE energy function is constantly good-performing when used to evaluate “partial” decoys. Second, both local item and global item should be normalized using match state size. Third, we prove that global item of the our scoring function is effective to capture distance interaction comparing with contact-preference based scoring functions. Fourth, we show that optimal local score can be used to determine “easy” pairs for which local score item is sufficient while adding global item may lead noise contrarily. Last, we train ω_{ L } on SALIGN [18] benchmark dataset.
Performance of dDFIRE on partial structure
Since we calculate dDFIRE energy on the “partial” decoy instead of a full-length structure, thus it is necessary to verify whether the “partial dDFIRE energy” still have the power to distinguish native-like decoys. To verify this, we performed experiments on three commonly-used benchmark datasets: LKF [1], Gapless Threading [2] and Rosetta [5]. The datasets contain 178, 200 and 232 proteins, respectively; and for each protein, 100 decoys were generated as control to the native structure. The objective of this experiment is to verify whether the “partial” native structure can be distinguished from the “partial” decoys by dDFIRE.
Score normalization
Effect of global items
We further investigate the effect of global item. As control, we performed comparison with the traditional way to describe non-local interactions via contact preference matrix [13, 15], i.e, S_{ p } = ∑_{ i }∑_{ j }δ(i, j)Pair(A(i), A(j)), and where A(i) is the matched residue in the sequence, δ(i, j) indicates whether i th and j th residue in the template have contact, P_{ m } is the profile vector at the m th position and C is the contact preference matrix.
We first give some notations before presenting the experiments to examine the effects of global items. For each query-template pair, two typical alignments are generated: the structural alignment A_{ R } generated via running TMalign [19], and the optimal alignment (denoted as A_{ L }) when only local item L(A) is taken into consideration, i.e. A_{ L } = argmin_{ A }L(A). For each alignment A, its real quality is measured by TMscore [20], denoted as TM(A). We also use L(A) and G(A) to denote the local score and global score of A, and use C(A) to denote the contact-preference-based score of A.
Determining pairs for which local score item alone is sufficient
On 144 of 200 pairs of SALGIN benchmark, local score alone is sufficient to find out a “good” alignment(TM(A_{ R }) – TM(A_{ L }) < 0.1). In fact, in these cases, adding global item may lead to false-negative [21]. We observed that the normalization of local scores help recognizing these “easy” pairs. This is reasonable since local score contains most of the homologous information between the sequence and the template.
Weight training process
Parameter ω_{ L } is trained by classification. For each query-template pair in SALIGN benchmark, one positive alignment A_{ p } and 10 negative alignments A_{ n } are selected (We also have tried other number of negative alignments, similar result is obtained).
Here, we use the reference alignment as positive alignment, i.e. A_{ p } = A_{ R }. Negative alignments are chosen from the top 100 alignments returned by dynamic programming. We first cluster these alignments to remove redundancy, and then randomly select alignments satisfying TM(A_{ p }) – TM(A_{ n }) > 0.2.
ω_{ L } should divide A_{ n } and A_{ p } as much as possible. Formally
ω_{ L } = arg max |H|
where
H = {(A_{ p }, A_{ n }) ∈ P|ω_{ L }L(A_{ p }) + G(A_{ p }) <ω_{ L }L(A_{ n }) + G(A_{ n })}
= {(A_{ p }, A_{ n }) ∈ P|ω_{ L }L(A_{ p }) – L(A_{ n })) <G(A_{ n }) – G(A_{ p })}.
After obtaining the parameter ω_{ L } and θ, our threading algorithm can be described informally as follows: given a query-template pair, dynamic programming algorithm is employed to calculate A_{ L } . If L(A_{ L }) <θ, then A_{ L } is considered as a good alignment and returned. Otherwise, local search algorithm is then used to find a better alignment under scoring function score(A) = ω_{ L }L(A) + G(A). The initial alignments used in this step are chosen from the dynamic programming table in the previous step.
Preliminary results on alignment generating
We test our threading method on Prosup benchmark (containing 127 query-template pairs). Each query-template pair shares low sequence identity but high structure similarity. Denote the alignment generated by our method as A_{ O }.
Alignment Accuracy Comparison on Prosup Benchmark.
Methods | Accurate(%) | ±4-residues(%) |
---|---|---|
FASTA | 31.4 | - |
Sequence | 34.1 | - |
PSI-BLAST | 35.6 | - |
RAPTOR | 44.0 | 63.7 |
A L | 43.1 | 63.0 |
A O | 52.0 | 70.1 |
Methods
Threading Algorithm
The framework of our threading algorithm are described as follows:
Algorithm 1(Threading Algorithm)
Input: query sequence, template, θ, ω_{ L }, α, k
Output: an alignment between query and template
step 1 set score(A) = L(A), calculate the optimal alignment A_{ L } under this scoring function by dynamic programming algorithm, save the best 100 alignments from the dynamic programming table
step 2 calculate L(A_{ L }) , if L(A_{ L }) <θ, then return A_{ L }
step 3 set score(A) = ω_{ L }L(A) + G(A), for each alignment A_{ i } in the 100 candidates in step 1, run local search algorithm(Algorithm 2 described in the following subsection) with parameter α, k and initial alignment A_{ i }, it returned A_{ Oi }
Local Search Algorithm
In this sub-section we describe the threading problem in a concise way, propose a local search algorithm based on a new neighborhood for general scoring function. Under a certain assumption, we prove its approximation guarantee for two specific scoring functions.
Problem Formulation
We first give some formal definitions.
Definition 1. Given a template T = {t_{1}, t_{2}, ⋯, t_{ m }}, t_{ i } <t_{i+1}and a sequence S = {s_{0}, s_{1}, s_{2}, ⋯, s_{ n }}, s_{ i } <s_{ i }_{+1}, a valid alignment is a non-decreasing mapping A from T to S.
Denote all valid alignments as F . Non-decreasing mapping is equivalent with traditional alignment definition with gap. For all t satisfying A(t) = s, s >s_{0}, we can define the smallest t actually matches with s while others are gap on template. In order to allow gap on the left end of the template, we add a extra amino acid s_{0} in the left end of the sequence. All t ∈ T aligned to s_{0} are gap on the left end of the template. Mapping allows gap on sequence naturally.
Now we define the neighborhood of an alignment. Denote the k- neighbor of A as N(A, k), we have the following definition.
Definition 2. Suppose A′ ∈ F, then A′ ∈ N(A, k) if and only if there exists a subset U of S, satisfying |U| ≤ k and ∀t ∈ T A′(t) ∈ {A(t)} ∪ U.
Intuitively, a member of k-neighbors of A differs with A only on at most k positions at sequences.
Claim 1. N(A, k) ⊂ N(A, k + 1).
Claim 2. N(A, n + 1) = F .
Claim 3. |N(A, k)| = O(m^{2}^{ k }n^{ k }), ∀A ∈ F .
Claim 1 and claim 2 are obvious. The proof claim 3 is put in the Appendix.
Claim 1 and claim 2 show that with the increasing of k, the number of neighbors of an alignment is growing and eventually reaches the whole space. Claim 3 estimates the size of |N(A, k)|. It shows that for a fixed k the number of neighbors of a valid alignment is polynomial about m and n.
Definition 3. For any A ∈ F, there is a real positive number denoted as score(A) to evaluate A, the threading problem is min_{ A }_{∈}_{ F }score(A).
score(A) is the general representation of scoring function. In this study, score(A) = ω_{ L }L(A) + G(A).
Algorithm
Based on the definition of neighborhood above, we give the local search algorithm as follows:
Algorithm 2(Local Search)
Input α ≥ 0, k, initial alignment A_{0}
Output an approximate local optimal solution of the scoring function score(A)
step 1 i = 0, initialize A_{0} according to input
step 2 calculate A_{ i }_{+1} = argmin_{ A }_{∈}_{ N }_{(}_{ A }_{ i }_{,}_{ k }_{)}score(A)
step 3 if (1 + α)score(A_{ i }_{+1}) <score(A_{ i }), i = i + 1, goto step 2
step 4 output A_{ i }
When α = 0, we can obtain an accurate local optimal solution. When α = 0 and k = n + 1, we can obtain an accurate global optimal solution.
Claim 4. Suppose α >0, The time complexity of algorithm 2 is O(m^{2}^{ k }n_{ k }log_{1+}_{ α }M), where
Proof. Based on the algorithm, we have
(1 + α)^{ i }score(A_{ i }) <score(A_{0}),
According to claim 3, each iteration wastes at most O(m^{2}^{ k }n^{ k }) time, so the claim is proved.
If a closing assumption is satisfied, we can prove two approximation guarantee results when the scoring function only consists of local item and pairwise contact item. Details are listed in the Appendix section.
Discussion
In order to employ general energy function, the key step is transforming alignment to decoy efficiently In this paper, “partial“ decoy strategy is quick enough but not accurate because only matched residues’ backbone and C_{ β } atoms are kept. Methods that effectively recover other unmatched residues and even side chain atoms according to alignments are imperative.
Though the energy function of Ab Initio can be used by threading, the two methods have fundamental difference on the divergence of search space. Actually, the search space of threading is much smaller than that of Ab Initio methods because many useful prior knowledge can greatly narrow its search space. For instance, we can restrict that a core on template either totally aligned or totally gaped. This prior has been verified and applied by many threading methods. Consequently, the search space can be reduced to O(N^{ m }) where N is the number of cores and m the length of query sequence. On average, N ≤ 10, it is much smaller than 200^{m}, which is the search space of ROSETTA.
In this paper we have proposed a local search algorithm to find out the optimal solution of general scoring function. This algorithm is based on a neighborhood definition, and this neighborhood can also be used by other search strategies such as simulated annealing and genetic algorithm.
Declarations
Acknowledgements
This article has been published as part of BMC Bioinformatics Volume 12 Supplement 1, 2011: Selected articles from the Ninth Asia Pacific Bioinformatics Conference (APBC 2011). The full contents of the supplement are available online at http://www.biomedcentral.com/1471-2105/12?issue=S1.
Authors’ Affiliations
References
- Loose C, Klepeis J, Floudas C: A new pair-wise folding potential based on improved decoy generation and side-chain packing. Proteins: Structure, Function, and Bioinformatics 2004, 54(2):303–314.View ArticleGoogle Scholar
- Zhang J, Chen R, Liang J: Empirical potential function for simplified protein models: Combining contact and local sequence–structure descriptors. Proteins: Structure, Function, and Bioinformatics 2006, 63(4):949–960.View ArticleGoogle Scholar
- Ranjit B, Pinak C: Discriminating the native structure from decoys using scoring functions based on the residue packing in globular proteins. BMC Structural Biology 9:Google Scholar
- Shen M, Sali A: Statistical potential for assessment and prediction of protein structures. Protein Science 2006, 15(11):2507–2524.PubMed CentralView ArticlePubMedGoogle Scholar
- Simons K, Kooperberg C, Huang E, Baker D: Assembly of protein tertiary structures from fragments with similar local sequences using simulated annealing and bayesian scoring functions1. Journal of Molecular Biology 1997, 268: 209–225.View ArticlePubMedGoogle Scholar
- Zhou H, Zhou Y: Distance-scaled, finite ideal-gas reference state improves structure-derived potentials of mean force for structure selection and stability prediction. Protein Science 2002, 11(11):2714–2726.PubMed CentralView ArticlePubMedGoogle Scholar
- Pearson W, Lipman D: Improved tools for biological sequence comparison. Proceedings of the National Academy of Sciences of the United States of America 1988, 85(8):2444.PubMed CentralView ArticlePubMedGoogle Scholar
- Altschul S, Gish W, Miller W, Myers E, Lipman D: Basic local alignment search tool. Journal of molecular biology 1990, 215(3):403–410.View ArticlePubMedGoogle Scholar
- Altschul S, Madden T, Schaffer A, Zhang J, Zhang Z, Miller W, Lipman D: Gapped BLAST and PSI-BLAST: a new generation of protein database search programs. Nucleic acids research 1997, 25(17):3389.PubMed CentralView ArticlePubMedGoogle Scholar
- Durbin R: Biological sequence analysis: probabilistic models of proteins and nucleic acids. Cambridge Univ Pr 1998.Google Scholar
- Soding J: Protein homology detection by HMM–HMM comparison. Bioinformatics 2005, 21(7):951.View ArticlePubMedGoogle Scholar
- Xu Y, Xu D, Uberbacher E: An Efficient Computational Method for Globally Optimal Threading1. Journal of Computational Biology 1998, 5(3):597–614.View ArticlePubMedGoogle Scholar
- Xu J, Li M, Kim D, Xu Y: RAPTOR: optimal protein threading by linear programming. INTERNATIONAL JOURNAL OF BIOINFORMATICS AND COMPUTATIONAL BIOLOGY 2003, 1: 95–118.View ArticleGoogle Scholar
- Lathrop R: The protein threading problem with sequence amino acid interaction preferences is NP-complete. Protein Engineering Design and Selection 1994, 7(9):1059.View ArticleGoogle Scholar
- Xu Y, Xu D: Protein threading using PROSPECT: design and evaluation. Proteins: Structure, Function, and Bioinformatics 2000, 40(3):343–354.View ArticleGoogle Scholar
- Wang S, Zheng W: CLePAPS: fast pair alignment of protein structures based on conformational letters. Journal of bioinformatics and computational biology 2008, 6(2):347–366.View ArticlePubMedGoogle Scholar
- Eswar N, Webb B, Marti-Renom M, Madhusudhan M, Eramian D, Shen M, Pieper U, Sali A: Comparative protein structure modeling using MODELLER. Curr Protoc Protein Sci 2007.Google Scholar
- Marti-Renom M, Madhusudhan M, Sali A: Alignment of protein sequences by their profiles. Protein Science 2004, 13(4):1071–1087.PubMed CentralView ArticlePubMedGoogle Scholar
- Zhang Y, Skolnick J: TM-align: a protein structure alignment algorithm based on the TM-score. Nucleic acids research 2005, 33(7):2302.PubMed CentralView ArticlePubMedGoogle Scholar
- Zhang Y, Skolnick J: Scoring function for automated assessment of protein structure template quality. Proteins: Structure, Function, and Bioinformatics 2004, 57(4):702–710.View ArticleGoogle Scholar
- Peng J, Xu J: Boosting protein threading accuracy. In Research in Computational Molecular Biology. Springer; 2009:31–45.View ArticleGoogle Scholar
- Domingues F, Lackner P, Andreeva A, Sippl M: Structure-based evaluation of sequence comparison and fold recognition alignment accuracy1. Journal of molecular biology 2000, 297(4):1003–1013.View ArticlePubMedGoogle Scholar
- Zhou H, Zhou Y: Fold recognition by combining sequence profiles derived from evolution and from depth-dependent structural alignment of fragments. Proteins: Structure, Function, and Bioinformatics 2005, 58(2):321–328.View ArticleGoogle Scholar
Copyright
This article is published under license to 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.