Fragment assembly is a powerful method of protein structure prediction that builds protein models from a pool of candidate fragments taken from known structures. Stochastic sampling is subsequently used to refine the models. The structures are first represented as coarse-grained models and then as all-atom models for computational efficiency. Many models have to be generated independently due to the stochastic nature of the sampling methods used to search for the global minimum in a complex energy landscape. In this paper we present , a fragment-based approach which shares information between the generated models and steers the search towards native-like regions. A distribution over fragments is estimated from a pool of low energy all-atom models. This iteratively-refined distribution is used to guide the selection of fragments during the building of models for subsequent rounds of structure prediction. The use of an estimation of distribution algorithm enabled to reach lower energy levels and to generate a higher percentage of near-native models. uses an all-atom energy function and produces models with atomic resolution. We observed an improvement in energy-driven blind selection of models on a benchmark of in comparison with the AbInitioRelax protocol.
Citation: Simoncini D, Zhang KYJ (2013) Efficient Sampling in Fragment-Based Protein Structure Prediction Using an Estimation of Distribution Algorithm. PLoS ONE 8(7): e68954. https://doi.org/10.1371/journal.pone.0068954
Editor: Yang Zhang, University of Michigan, United States of America
Received: April 18, 2013; Accepted: June 7, 2013; Published: July 25, 2013
Copyright: © 2013 Simoncini, Zhang. 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: Funded by RIKEN Initiative Research Unit program. The funder had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
The prediction of protein structures from their sequences has been a subject of intense research ever since the seminal work of Anfinsen . Based on the principle that form follows function, the three-dimensional (3D) structure of a protein provides critical clues to its function. Moreover, the knowledge of a protein structure facilitates the design of therapeutic agents that modify its function. There are two main challenges that a protein structure prediction (PSP) method has to face: the inaccuracy in energy functions and the size of the search space. The inaccuracy in energy functions make identification of near-native models a difficult task. A study suggests that in some cases the native structure does not belong to the global minimum basin , and it was estimated that for over of the tested targets, a better sampling of the search space may lead to successful predictions. Inaccuracies in the energy functions combined with the huge size of the conformational space give rise to sampling issues. Many methods have been proposed to deal with these problems. One idea was to reduce the search space by assembling the structure from a pool of experimentally determined structural fragments . This fragment assembly approach has become one of the most popular methods for protein structure prediction due to the success of , , . employs a two stage strategy: fragment assembly followed by all-atom refinement. During the fragment assembly, the protein models are represented by backbone atoms and centroid of side chains (coarse-grained sampling). Once the models are assembled, the structure representation is switched to all-atom: side chains are added and packed by minimizing an all-atom knowledge-based energy function .
Many other fragment-based approaches have been proposed. The method, which was successful in recent CASP experiments , uses variable length fragments and replica exchange Monte Carlo for sampling . attempts to improve the conformational sampling efficiency by using Conformational Space Annealing . introduces the concept of reversible fragment insertion, where local structures created by the junction of two proteins fragments can be reused later on during the sampling process . combines variable length fragments and fold recognition analysis. It uses a genetic algorithm for sampling . combines supersecondary structural fragments built from several sequential secondary structures with small fragments. The models are assembled using a genetic algorithm and simulated annealing . Some probabilistic methods estimate the joint angle distribution by a mixture model of particular distributions , . Among them, Fragment-HMM uses protein fragments to obtain a first estimation of the torsion angle distributions using the cosine model, a bi-variate von Mises distribution. The cosine models are used as hidden nodes in a position-specific hidden Markov Model which is then used to sample a sequence of torsion angle pairs. The method then uses the generated protein models as new input to refine the joint angle distributions and iterates until convergence. The incorporation of information from prior rounds has been shown to be beneficial for the prediction of protein secondary and tertiary structures , . Using the principle of sequential stabilization, Adhikari et al. has demonstrated that the accuracy of predicted structures can be greatly improved using a process of progressive learning and structural stabilization found in prior round of folding.
Fragment-based methods such as need to generate a huge number of models in order to find a correct structure. The rugged nature of the energy landscape, due to the inaccuracy of the energy functions and the size of the search space, necessitates the use of stochastic sampling methods. Recently, we have proposed a method () that takes advantage of the large size of the data-set by enabling communications between predictions during the sampling . The idea was that if some fragments occur more often in low energy models, then these fragments are more likely to resemble the native structure. Using an Estimation of Distribution Algorithm (EDA), iteratively updates the probabilities of inserting fragments in new models according to their frequency in low energy models. The results obtained with were promising, and show that the method is able to enhance the proportion of near-native structures in the pool of protein models on a benchmark of proteins. However, it employed coarse-grained models to estimate fragment probabilities. As a result, some high quality models could not be identified and the sampling was misguided on a few targets because of energy function inaccuracy. Also, as its aim was to study the sampling dynamics at a coarse-grained level, it lacks the ability to produce all-atom models.
In this paper, we describe a new method () that estimates probability mass functions of fragments from all-atom models instead of coarse-grained models as . Our new protocol relies on 's all-atom energy in order to rank the models during the estimation of distribution step of the algorithm and the all-atom models produced show notable improvements over and AbInitioRelax. The gain in accuracy provided by the all-atom energy function had several positive effects on the sampling; the closest structure to native in the pool of models, the proportion of near-native models and the accuracy of lowest energy models improved on average, and on a majority of the proteins in our benchmark.
favors fragments from the library that are closer to native fragments. The probability of selecting each of the 25 9-residue fragments at iteration 4 of is plotted against the CαRMSD of each fragment to the native structure for 3 fragment windows of PDB codes 1ogw, 1dtj and 1bq9 in Figure 1. Pearson product moment correlation coefficient shows anti-correlation between probabilities and CαRMSD to native, putting in evidence that native-like fragments usually get a high probability of being selected and vice versa. The average CαRMSD of fragments to native weighted by their probability of being selected at iterations 1 and 4 is plotted for the same PDB codes in Figure 2. Overall, the average probability weighted CαRMSD shows improvement from iteration 1 to 4 in all of the 3 cases.
The Pearson correlation coefficient (R) between probabilities and CαRMSD to native structure is given for each fragment window.
The average over all fragments windows at iterations 1 and 4 shows an overall improvement on all cases: PDB codes 1ogw, 1dtj and 1bq9.
can successively generate lower energy models after each iteration. The distributions of the energies of the models generated by at iterations and and by were shown in Figure 3. The measures were made with two protein targets, PDB codes 1bq9 and 1ogw. Whereas the distributions of 's iteration and have the same shape and suggest that the two methods sample similar regions of the search space and in equal proportions, the curve describing the iteration is different. In both cases, favors the sampling of low energy basins of the search space. For 1bq9, even though the shape of the curves are the same, we note that the curve at iteration from already shifts a little towards lower energies.
The estimation of distribution algorithm allows to increase the performance from iteration to iteration .
The improvement in lower energy basins sampling achieved by translates into better quality models with lower CαRMSD (C-alpha Root Mean Square Deviation) to native. This can be seen from Figure 4, which plots the distribution of decoys as a function of CαRMSD to native. The results match with the observations made at energy level. At first iteration, sample similar regions of the search space. As a shift at iteration for 1bq9 was observed, it can be seen in this figure that produces models closer to the native structure even at iteration . The distributions change at iteration . In both cases, regions closer to the native structure are thoroughly sampled at this stage. The algorithm is particularly efficient for 1bq9. This efficiency of the algorithm could be due to the shape of the fitness landscape induced by all-atom energy. Since favors the insertion of fragments which have been identified as being helpful in minimizing the energy, a good correlation between the lowest energies and CαRMSD to native is one reason for improved performance.
Whereas the distributions look identical between iteration and , at iteration the distribution has shifted towards native structure.
The quality of models have been improved by estimating probability mass functions of fragments based on all-atom energy instead of coarse-grained energy. The distribution of energies as a function of CαRMSD to native for all models generated at iteration and produced by for one target, 1bq9, are shown in Figure 5. In addition, it also shows the scatter plot between energy and CαRMSD to native for the same target with which was using 's coarse-grained energy. This figure revealed the critical importance of the energy function in our process. In , even though some good models were discovered, the sampling process was misled by the inaccuracies of the coarse-grained energy function. As a result, enhanced the search in the wrong region at about 9 Å from the native structure. The high number of low energy misfolded models complicates the identification of the high quality ones. Once we use the all-atom energy in our process, the fitness landscape changes. This time, the algorithm can discover unprecedented energy levels which correspond to models located at about 1 Å CαRMSD from the native structure. This suggests that using a more accurate energy function for the periodic estimation of distributions in improves the sampling.
We know that turning coarse-grained models into all-atom models dramatically modifies the landscape, but it is unclear if using the all-atom energy to drive 's sampling is more efficient. We performed some control experiments for which we used to generate coarse-grained models, and turned them into all-atom models with 's fast Relax protocol as a final step. The results were then compared with those from our new algorithm which includes fast Relax at each iterative step. The same number of models was produced with each method for protein sequences and each protocol's ability to produce near-native models was examined. Table 1 shows that systematically produced models with smaller CαRMSD to the native structure, and also generated a higher proportion of near-native models than . This result shows that switching the models to an all-atom representation at each iteration of our algorithm and modifying the probabilities of subsequent fragment selection according to 's all-atom energy function improves the accuracy of .
can generate a higher proportion of near-native models on average. The Table 2 shows a comparison of the CαRMSD to native that each method can reach. The average lowest 1 CαRMSD and 1% CαRMSD to native are shown. In addition, the CαRMSD to native of the best model is shown. The analysis of the proportion and quality of near-native models produced by and shows that can generate a higher proportion of near-native models on average. The best model generated by each method without selection considerations is also slightly better on average for . Table S1 shows that the trend is similar when looking at AARMSD (All-Atom Root Mean Square Deviation) values. The trend is the same whether we compare CαRMSD or AARMSD.
The improvement of low quality models away from native is probably not very meaningful. Models with CαRMSD less than 3 Å from native structure could potentially be used as templates for solving crystal structures by molecular replacement . A global view of the ability of each method to produce and identify near-native models was given in Figure 6. The percentage of models within 3 Å CαRMSD of the native structure for the lowest energies were computed. For each target, the difference between the percentage of near-native models obtained from 's and 's dataset was plotted. This difference is in favor of for all points above the straight line. It is in favor of for all points under it. A majority of the points are above the straight line, which confirms that improves the quality of the lowest energy models. The percentage of models generated closer than 3 Å CαRMSD from the native structure amongst the lowest energies in 's and 's datasets was analyzed in details and illustrated in Figure S1. Measurements range from models within Å to less than Å from native by steps of Å. was able to produce a higher percentage of models at less than Å for targets. outperforms on targets. Neither of the two methods was able to generate near-native models within the lowest energies on the same two targets: PDB codes and .
The percentage of models less than 3 Å away from native in terms of CαRMSD was computed. The differences in percentages between and are plotted.
The improved distribution of low energy models has enabled the “blind selection” of better models amongst all generated by . The “blind selection” results with energy as a selection criterion for and are shown in Table 3. Two results, the first and best predictions, are presented. The first prediction is the model with the lowest energy in the dataset whereas the best prediction is the best model out of the lowest energies. Both first and best prediction of improves over of about Å either when looking at CαRMSD or AARMSD (See Table S2 for AARMSD values).
Finally, the improved “blind selection” of models is due to the iterative estimation of distribution over the fragments. A comparison of the blind selection ability of at iterations and is given in Table 4. The first and best of predictions are shown for models taken from iteration alone and iteration alone. As stated in the method section, produces one quarter of the total number of models at each iteration. The results show an improvement of Å on average for the first prediction and over Å on average for the best prediction. For comparison, one quarter of the final models generated by AbInitioRelax was randomly selected. The results show that performs slightly better than at iteration , when no information is shared to produce the models. This result suggests that the sampling algorithm of is more efficient than ours. The same trend was observed when looking at AARMSD values (see Table S3). Nevertheless, outperforms on this benchmark thanks to the use of EDA. Fragment-based approaches appear to gain considerable advantage from sharing information between predictions.
The identification of near-native models is challenging due to the inaccuracy of the energy function. In our previous communication, we presented which focused on enrichment of the near-native structures at a coarse-grained level. We have demonstrated that the improved coarse-grained models can lead to better all-atom models by refining the top quality coarse-grained models into all-atom models . However, the question of how top quality coarse-grained models are identified was not addressed. We were focused on the generation rather than the identification of good models, believing that it is pointless trying to identify good models if they are not generated to begin with. Here we show that it is not necessary to identify good coarse-grained models for refinement into all-atom models. Our new method takes all of the coarse-grained models and refines them into all-atom models. By interlacing all-atom representation of models and coarse-grained sampling, is able to enrich the proportion of near-native structures in the lowest energy all-atom models. Therefore, we have demonstrated here that the improved coarse-grained models can lead to better all-atom models without the need to identify good coarse-grained models.
uses the lowest energy all-atom models for the estimation of fragment distributions, whereas uses the lowest energy coarse-grained models for the estimation of fragment distributions. The estimated fragment distributions are used for the assembly of coarse-grained models in both methods. It is generally considered that the all-atom energy function is more accurate than the coarse-grained energy function. The all-atom models should be better than coarse-grained models given that the former has the benefit of complete side-chains taken into consideration. Therefore, it makes sense to derive probability mass functions of fragments from the refined all-atom models, and then to use these probabilities to guide the assembly of fragments. We have found that quality of coarse-grained models generated by has been improved over those generated by for the same target. By using this strategy, we benefit from the speed of the coarse-grained energy function during the sampling and of the accuracy of the all-atom energy which is helpful to guide the search in subsequent coarse-grained sampling rounds. Our results show the efficiency of this protocol for blind selection of models: at iteration the quality of the lowest energy model and of the closest structure to native out of the top lowest energy models dramatically improves over iteration .
The fragment assembly approach requires the generation of a huge number of models in order to produce a satisfactory solution. This is due to the stochastic nature of the sampling methods and to the inaccuracy of knowledge-based energy functions. The predictions are typically independent and modern technology allows massive parallelization of the computation. In this paper, we show that sharing information between these independent predictions can improve the quality of final results. The estimation of distribution over the fragment library relying on each fragment's frequency in low energy models allows to significantly improve its performance in iterations. The amount of communications remains small, and the effective parallelization ratio is around . Our study suggests that the gain in performances is independent of the sampling method. Therefore, estimation of distribution can possibly increase the efficiency of any fragment-based approach.
The computation time required for the implementation of the estimation of distribution algorithm, including the encoding and decoding of fragments, the calculation of the probability mass functions and the sampling of fragments with Roulette wheels, is a small fraction of the time used for the generation of each model. In a previous communication, we reported that was 2.5 times slower than . This was derived from the comparison of the times it took for the generation of equal numbers of coarse-grained models by both methods. The difference was due to our implementation of the simulated annealing and iterated hill climbing protocols in the coarse-grained model generation step is less efficient than the Monte Carlo search protocol in Rosetta. However, the extra time that spent on the coarse-grained model generation did not produce better results than when EDA was not used as shown in Table 4. The quality of all-atom models from “iteration 1” without EDA was comparable or slightly worse than that of for both first prediction and best prediction. It is only when EDA was used that the all-atom models were improved as shown from “iteration 4” in Table 4 compared to as well as “iteration 1”. Since our goal is to evaluate the impact of EDA on the all-atom models due to the improved coarse-grained model sampling and the computing time required for the implementation of EDA is negligibly small, the same number of all-atoms models were generated when comparing the performance of and in this paper.
The incorporation of information from prior rounds in an iterative process has been shown previously to be a powerful technique applicable to protein structure prediction. The principle of sequential stabilization applied to protein folding is such an example . The method uses the statistics of folding trajectories garnered from prior rounds to bias subsequent sampling of backbone dihedral angles, tertiary contacts and hydrogen bonds. There were no fragments used in the method.
Even though Fragment-HMM and both use iterative strategies and are similar in the sense that they use information on generated models to refine estimations, the two methods differ in many ways. First, whereas Fragment-HMM estimates a distribution over the torsion angles assuming it follows a bi-variate von Mises distribution, estimates a distribution over the fragment library starting from a uniform distribution. Then, the sampling method is different: Fragment-HMM uses a position specific hidden Markov Model and uses simulated annealing and iterated hill climbing. Also, even though Fragment-HMM initially uses fragments to estimate distributions at the first iteration, it doesn't use fragments during the sampling, unlike . Finally, Fragment-HMM iterates until converging on one final model, whereas generates a diverse set of models, out of which the best ones will be selected.
The competition between global and local interactions plays a critical role in protein folding as well as structure prediction. This has been exploited for improving protein secondary and tertiary structure predictions using the principle of sequential stabilization , . The identification of good quality fragments using the estimation of distribution algorithms implemented in considers global interactions since the distributions are derived from the lowest energy all-atom models, whereas the initial fragment library is obtained by sequence homology that only takes into account local interactions of the residues within the fragment window. As the quality of all-atom models improve after each iteration, the global interactions are more accurately represented, which enables the good quality fragments being identified. However, unlike which can identify novel interactions as they have been generated and bias sampling towards those novel interactions, seeks only to bias existing fragments in the library, although novel fragments have been generated in the structure prediction process. The identification of novel fragments and bias the sampling towards good quality novel fragments will be an interesting direction for future exploration.
Our study also showed some deficiencies in 's sampling method. It fails to perform as well as when no information is shared between the models (i.e. at iteration ). This observation leaves room for improvement of our method. Beyond the improvement of our heuristic sampling methods, future work will focus on other ways of sharing information between predictions.
is a fragment-based protein structure prediction algorithm. Similarly to , it has two stages. First, 9-mers (followed by 3-mers) are assembled together to create coarse-grained models. 9-mers and 3-mers are taken from a fragment library which is created from protein structures available in the PDB. The fragment library we used was constructed using 's fragment picking method . When creating this library, proteins that shared more than 30% sequence identity with the target sequence were removed in order to remove any favorable bias. During the second stage, models are represented in atomic detail, and side chains are packed to minimize an all-atom energy function. The Relax protocol was used to perform this operation.
's protocol is described in Table 5 (Algorithm 1). It is an iterative algorithm using the concept of EDA to gather information between initially independent predictions. The concept of EDA is used to influence the probability of selecting fragments from the library. A fraction of the final model set (25%) is generated at each iteration. At iteration , there is a uniform distribution over the library and every fragment has the same probability of being selected. At each iteration, a fraction of the lowest energy models (10%) is selected as a sample set. To compute the energy, all the models are relaxed in their all-atom representation via Relax protocol. We keep track of the links between coarse-grained and all-atom representations of a model so that we can retrieve which fragments were used to generate which all-atom model. The probabilities of selecting fragments for insertion during subsequent iterations are modified according to the observed distribution of fragments used in the sample set. The sampling is influenced by the probability mass function defined over the fragment library. Still, each iteration starts with models in extended conformation. Models generated at a given iteration are stored in the final set and are not reused for subsequent iterations. inherits its sampling engine from . The sampling is performed using an alternation of simulated annealing  and iterated hill climbing . The estimation of distribution is handled by the function and is computed by the following formula:where is the current iteration, the probability of fragment , the observed frequency of fragment and a conservation rate (k = 0.6 in our experiments).
The performance of was measured on a dataset of protein sequences, which were used in our previous studies . We performed iterations of for each target. The models used during the estimation phase (iterations , and ) are part of the final model set: of the final models are generated at each iteration. The number of models generated (same for Rosetta and ) depends on the length of the target sequence: 250,000 for PDB codes 1m6t, 1iib, 1acf, 3chy and 300,000 for all other targets. CαRMSD calculations were performed with the ranker tool from Durandal . All-atom RMSD, referred to as AARMSD in the following, were computed with the LSQKAB program from the CCP4 Software Suite , . Version 3.2 was used for performance comparisons.
We present an estimation of distribution-based protein structure prediction algorithm which generates models with atomic details. The use of 's fast Relax protocol in the iterative process of allows the estimation of distributions over the protein fragment libraries according to an all-atom energy function. Energy distribution and CαRMSD to native histograms revealed that our protocol can reach lower energies and generate more accurate models after iterations. A comparison with AbInitioRelax shows that is able to produce more accurate models and a higher percentage of near-native structures. The proportion of near-native structures in the low energy range also improves on a majority of targets. As a result, energy-driven blind selection of models is more efficient: selects more accurate models when looking at the lowest or the top lowest energies in our dataset on a benchmark of protein targets.
is released under the GNU General Public License. It can be downloaded from http://www.riken.jp/zhangiru/software.html.
Models distribution as a function of CαRMSD to native for the lowest 500 energies in and datasets.
Comparison of all-atom models generated by and . 1‰ AARMSD is the average over the 1‰ lowest AARMSD to native models. Similarly, 1% AARMSD is the average over the 1% lowest AARMSD to native models. Best model is the single lowest AARMSD to native model. All mean differences are statistically significant with a confidence greater than 95% according to the Student's t-test.
Comparison of best all-atom models selected based on energy. The first prediction is the model with the lowest energy. The best prediction is the best model out of the five lowest energies. All results are shown as AARMSD to native structure.
Blind selection ability of all-atom models generated by at iterations and . The first prediction is the model with the lowest energy. The best prediction is the best model out of the five lowest energies. All results are shown as AARMSD to native structure. Models produced at iteration alone and iteration alone are compared. For comparison, the columns show the same data obtained from a sample of models randomly picked from 's prediction results.
We thank RIKEN, Japan, for an allocation of computing resources on the RIKEN Integrated Cluster of Clusters (RICC) system. We are grateful to the Rosetta Commons Member Institutions (http://www.rosettacommons.org/) for making Rosetta source code available. We thank Francois Berenger for providing efficient computational tools for statistical analysis. We thank Prof. Jeremy Tame for proofreading the manuscript.
Conceived and designed the experiments: DS KYJZ. Performed the experiments: DS. Analyzed the data: DS KYJZ. Wrote the paper: DS KYJZ.
- 1. Anfinsen CB, Haber E, Sela M, White FH (1961) The kinetics of formation of native ribonuclease during oxidation of the reduced polypeptide chain. Proc Natl Acad Sci U S A 47: 1309–14.
- 2. Kim DE, Blum B, Bradley P, Baker D (2009) Sampling bottlenecks in de novo protein structure prediction. J Mol Biol 393: 249–60.
- 3. Bowie JU, Eisenberg D (1994) An evolutionary approach to folding small alpha-helical proteins that uses sequence information and an empirical guiding fitness function. Proc Natl Acad Sci U S A 91: 4436–40.
- 4. Simons KT, Kooperberg C, Huang E, Baker D (1997) Assembly of protein tertiary structures from fragments with similar local sequences using simulated annealing and Bayesian scoring functions. J Mol Biol 268: 209–25.
- 5. Simons KT, Bonneau R, Ruczinski I, Baker D (1999) Ab initio protein structure prediction of CASP III targets using ROSETTA. Proteins Suppl 3: 171–6.
- 6. Rohl CA, Strauss CEM, Misura KMS, Baker D (2004) Protein structure prediction using Rosetta. Methods Enzymol 383: 66–93.
- 7. Leaver-Fay A, Tyka M, Lewis SM, Lange OF, Thompson J, et al. (2011) ROSETTA3: an objectoriented software suite for the simulation and design of macromolecules. Methods Enzymol 487: 545–74.
- 8. Kinch L, Yong Shi S, Cong Q, Cheng H, Liao Y, et al. (2011) CASP9 assessment of free modeling target predictions. Proteins: Structure, Function, and Bioinformatics 79: 59–73.
- 9. Xu D, Zhang Y (2012) Ab initio protein structure assembly using continuous structure fragments and optimized knowledge-based force field. Proteins 80: 1715–1735.
- 10. Lee J, Kim SY, Joo K, Kim I, Lee J (2004) Prediction of protein tertiary structure using PROFESY, a novel method based on fragment assembly and conformational space annealing. Proteins 56: 704–14.
- 11. Chikenji G, Fujitsuka Y, Takada S (2003) A reversible fragment assembly method for de novo protein structure prediction. The Journal of Chemical Physics 119: 6895–6903.
- 12. Karplus K, Karchin R, Draper J, Casper J, Mandel-Gutfreund Y, et al. (2003) Combining localstructure, fold-recognition, and new fold methods for protein structure prediction. Proteins 53 Suppl 6: 491–6.
- 13. Jones DT, McGuffin LJ (2003) Assembling novel protein folds from super-secondary structural fragments. Proteins 53 Suppl 6: 480–5.
- 14. Hamelryck T, Kent JT, Krogh A (2006) Sampling realistic protein conformations using local structural bias. PLoS Comput Biol 2: e131.
- 15. Li SC, Bu D, Xu J, Li M (2008) Fragment-HMM: a new approach to protein structure prediction. Protein Sci 17: 1925–34.
- 16. DeBartolo J, Colubri A, Jha AK, Fitzgerald JE, Freed KF, et al. (2009) Mimicking the folding pathway to improve homology-free protein structure prediction. Proc Natl Acad Sci U S A 106: 3734–3739.
- 17. Adhikari AN, Freed KF, Sosnick TR (2012) De novo prediction of protein folding pathways and structure using the principle of sequential stabilization. Proc Natl Acad Sci U S A 109: 17442–17447.
- 18. Simoncini D, Berenger F, Shrestha R, Zhang KYJ (2012) A probabilistic fragment-based protein structure prediction algorithm. PLoS One 7: e38799.
- 19. Blow DM, Rossmann MG (1961) The single isomorphous replacement method. Acta Crystallographica 14: 1195–1202.
- 20. Gront D, Kulp DW, Vernon RM, Strauss CEM, Baker D (2011) Generalized fragment picking in Rosetta: design, protocols and applications. PLoS One 6: e23294.
- 21. Kirkpatrick S, Gelatt CD, Vecchi MP (1983) Optimization by simulated annealing. Science 220: 671–680.
- 22. Lourenco H, Martin O, Stutzle T (2001) Iterated Local Search. In “Handbook of Metaheuristics”, Ed F Glover and G Kochenberger, ISORMS 57, p 321–353 (2002), Kluwer.
- 23. Berenger F, Zhou Y, Shrestha R, Zhang KYJ (2011) Entropy-accelerated exact clustering of protein decoys. Bioinformatics 27: 939–45.
- 24. Winn MD, Ballard CC, Cowtan KD, Dodson EJ, Emsley P, et al. (2011) Overview of the CCP4 suite and current developments. Acta Crystallogr D Biol Crystallogr 67: 235–42.
- 25. Kabsch W (1976) A solution of the best rotation to relate two sets of vectors. Acta Crystallographica A32: 922–923.