GalaxyTBM: template-based modeling by building a reliable core and refining unreliable local regions
© Ko et al.; licensee BioMed Central Ltd. 2012
Received: 23 February 2012
Accepted: 7 August 2012
Published: 10 August 2012
Protein structures can be reliably predicted by template-based modeling (TBM) when experimental structures of homologous proteins are available. However, it is challenging to obtain structures more accurate than the single best templates by either combining information from multiple templates or by modeling regions that vary among templates or are not covered by any templates.
We introduce GalaxyTBM, a new TBM method in which the more reliable core region is modeled first from multiple templates and less reliable, variable local regions, such as loops or termini, are then detected and re-modeled by an ab initio method. This TBM method is based on “Seok-server,” which was tested in CASP9 and assessed to be amongst the top TBM servers. The accuracy of the initial core modeling is enhanced by focusing on more conserved regions in the multiple-template selection and multiple sequence alignment stages. Additional improvement is achieved by ab initio modeling of up to 3 unreliable local regions in the fixed framework of the core structure. Overall, GalaxyTBM reproduced the performance of Seok-server, with GalaxyTBM and Seok-server resulting in average GDT-TS of 68.1 and 68.4, respectively, when tested on 68 single-domain CASP9 TBM targets. For application to multi-domain proteins, GalaxyTBM must be combined with domain-splitting methods.
Application of GalaxyTBM to CASP9 targets demonstrates that accurate protein structure prediction is possible by use of a multiple-template-based approach, and ab initio modeling of variable regions can further enhance the model quality.
KeywordsProtein structure prediction Model refinement Loop modeling Terminus modeling
Three-dimensional protein structures provide invaluable insights into the molecular basis of protein functions, and such insights are essential for rational design of molecules regulating these functions. Nowadays, in an increasing number of cases, it has become possible to model protein structures with acceptable accuracy by employing much less effort than that required in experimental methods. Progress in computational protein structure prediction has been boosted by methodological improvements in the technique called template-based modeling (TBM), which uses experimental structures of homologous proteins as templates. As biological sequence and structure databases expand continuously, TBM is expected to become an even more promising tool for practical molecular biology, pharmaceutical chemistry, and protein engineering problems .
Template-based modeling, also called homology modeling or comparative modeling, generally consists of the following steps [1, 2]: (1) identification of homologous proteins with known structures to be used as templates; (2) alignment of the sequences of the target and templates; (3) creation of model structures from the alignment; and (4) refinement of the models. Contemporary methods usually treat each stage separately, and the full TBM procedure can therefore be established by combining methods for each of the above stages.
Despite recent progresses, there still remain challenges for each stage mentioned above. One of the important challenges is how to optimally combine information from multiple templates to build a single model when experimental structures of multiple homologues are available. Using multiple templates rather than a single template offers several obvious benefits: the possibility of including a better template increases, and the fraction of the target sequence covered by templates is extended [3–5]. In addition, different regions in template structures may be combined to produce a more accurate model . However, in practice, it is complicated to combine information from multiple templates in an optimal way . Since the average quality of multiple templates is bound to be worse than that of the single best template, using multiple templates is associated with a rather large risk of contaminating reliable information from the best template. To overcome this problem, various approaches have been proposed [1, 7, 8]. Most of them heavily rely on a single top template while additional templates are used to fill the gaps not covered by the top template [3, 9].
Another challenge is to model structurally variable regions among templates or those regions not covered by any templates, which we call ULRs (unreliable local regions). Unless the target sequence is quite similar to those of the templates (for example, with sequence identity > 30%), the expected quality of template-based models could be limited by such regions. Moreover, such ill-conserved regions where sequence insertion/deletion occurs may not be the subject of typical TBM. Despite previous efforts, progress in modeling such regions seems to be rather limited . Since high-resolution models are required for practical applications, it is doubtless that better ULR modeling is essential.
We recognize that the above 2 challenges are not independent of each other. For example, the performance of ULR modeling can be limited by the quality of the framework structure constructed from multiple templates [10, 11]. We therefore propose a strategy by which both initial TBM and subsequent ULR modeling can benefit from each other. In the initial TBM, we focus on accurate modeling of more conserved regions among multiple templates, without the need to consider potentially unreliable regions since such regions are taken care of in the ULR modeling stage. In the ULR modeling stage, we fix the more reliable core structure so as not to deteriorate the overall model quality by potentially less reliable ab initio ULR modeling. Therefore, ULRs can be modeled in a more accurate framework structure, and the conformational search space for ULR modeling is also effectively reduced to the local regions. Related approaches that construct a reliable core and refine unreliable regions have been proposed previously [12, 13]. The difference between our approach and these is that we put more stress on the “accuracy” (rather than the “coverage”) of the core structure in the initial TBM stage (See METHODS for details).
We call this new method GalaxyTBM, as it is based on the GALAXY molecular modeling package [11, 14–16]. GalaxyTBM employs a multiple-template method designed to produce reliable core structures by rescoring HHsearch  results for multiple-template selection and by core sequence alignment using PROMALS3D . Model building from the alignment and subsequent ULR modeling is performed using optimization modules in GALAXY [11, 16].
All components of the prediction pipeline were tested in the 9th critical assessment of techniques for protein structure prediction (CASP9) as a predictor named “Seok-server.” According to the official results of CASP9, Seok-server is recognized as one of the top 6 servers . Since the prediction strategy for Seok-server had to be modified a few times during CASP9, as the method was immature at the beginning, the most recent version, GalaxyTBM, is presented here. When GalaxyTBM was tested on 68 single-domain CASP9 TBM targets, fixing the structure database at the version with which Seok-server was used during CASP9, it reproduced the performance of Seok-server (average GDT-TS of 68.1, compared to 68.4 for Seok-server). Performance of the TBM pipeline was evaluated by analyzing the improvements achieved at each stage. Merits of the new components in the pipeline over other TBM methods are also discussed.
Results and discussion
Rescoring of HHsearch results improves the template quality
We used a simple but effective rescoring strategy to select multiple templates from the homologues detected by HHsearch , as described in the METHODS section. Here, we analyzed the performance of the rescoring method in terms of the quality of the top ranker compared to that of the HHsearch top ranker. Template quality was measured by a similarity score called TM-score calculated using the TM-align tool , which ranges from 0 (no similarity) to 1 (same as the native structure). Improvement achieved by the selection scheme of “multiple” templates is discussed in the next subsection.
Overall, top rankers obtained by the rescoring scheme were closer to the native structures of the target proteins than the HHsearch top rankers, when tested on the 68 single-domain CASP9 TBM targets. Different proteins ranked as top by rescoring in 19 out of the 68 cases, with an average improvement of 0.046 in TM-score. TM-score increased for 15 out of the 19 targets and decreased for the remaining 4 targets, with average increases of 0.072 and −0.033, respectively. A paired t-test for the TM-score changes for the 19 targets showed that the improvement is statistically significant, with a P-value of 0.0072.
Multiple-template information improves the model quality
By using the current multiple template approach, GDT-TS improved when compared to the values obtained using the single-template approach (the sum over 68 domains increased by 3.4% from 4429.3 to 4580.9 with an average improvement of 2.23 per domain) and the naïve multiple-template method (the sum increased by 3.8% from 4412.7 to 4580.9 with an average improvement of 2.47 per domain). The GDT-TS improvement is statistically significant, with P-values of 0.006 and 0.02, respectively. Improvement over the naïve method was prominent when high-ranking proteins by HHsearch have diverse structures, implying that the current multiple-template selection scheme that excludes dissimilar structures is fairly successful. For example, for T0539, mean GDT-TS of generated models is 75.59 and 59.14 by the current approach and by the naïve approach, respectively. Similar type of large GDT-TS improvement of > 5 was also found in T0532, T0552, T0559, T0614, and T0643.
To determine whether the model improvement by the multiple-template approach is a consequence of covering more residues by additional templates, we checked whether a similar improvement was found for core regions (Figure 2). The core region is defined here as the target residues aligned with the single top template by PROMALS3D . The core region covers 31% to 100% of the whole protein, with an average coverage of 85%. As shown in Figure 2A, GDT-TS of the core region was also improved by the current multiple-template method compared to the single-template method. Average GDT-TS improvement was 2.08%, with a P-value of 0.0106.
In conclusion, the current multiple template selection method contributes to improving the core structure by utilizing useful information from additional templates selected by the current pipeline.
Better optimization during model-building further improves the model quality
In GalaxyTBM, model building is performed by the MODELLER-CSA  module implemented in GALAXY. It was previously reported that more thorough optimization of the target restraint function derived by MODELLER is possible with the method, generating model structures more consistent with the restraints . To evaluate the performance of model building in the current pipeline, we compared the structures generated in this stage with the model structures generated simply by using MODELLER . The 2 methods, MODELLER and MODELLER-CSA, use the same sequence alignments, template lists, and therefore the same spatial restraints, and differ only in the optimization method.
As in the previous subsection, 100 model structures were generated for each target. Overall, model building by MODELLER-CSA improved the sum of GDT-TS by 0.6% (from 4580.9 to 4622.2) compared to MODELLER, with a P-value of 0.002. Average GDT-TS improvement was 0.13 for the 25 targets for which single templates were used and 0.87 for the 43 targets for which multiple templates were used. The better GDT-TS improvement in the multiple-template cases can be explained by the fact that the more complex target restraint functions for the multiple-template problems can be better optimized with the more rigorous optimization method .
In addition to the backbone structure quality, the side-chain structure quality was also improved by the better optimization during model building. The χ1 accuracy (percentage of the cases in which χ1 is within 30° from the native value) was improved in 65 out of 68 targets, with an average improvement of 5.9%. The χ1 + χ2 accuracy (percentage of the cases in which both χ1 and χ2 are within 30° from the native values) was also improved (improved in 63 out of 68 cases, with an average improvement of 4.5%). This improvement is consistent with the findings of the previous report by Joo et al..
Positive effects of the overall multiple-template strategy
The above analysis indicates that the positive effects of using multiple templates can be maximized by the current template selection strategy that considers core structure consensus and the more rigorous optimization during model-building, and the common adverse effects caused by including inconsistent templates in typical multiple-template methods can be minimized by use of such a combination.
ULR refinement also contributes to improvement of the model quality
Here we present the results of the final stage of the pipeline, i.e., refinement of ULRs. A total of 204 ULRs (56 termini and 148 loops) were detected in the initial stage, and 132 ULRs were finally subject to reconstruction following the selection rule described in METHODS. These reconstructed ULRs consisted not only of the regions that were not aligned to any template residues but also of the regions that were structurally inconsistent among templates.
Template selection and multiple sequence alignment take a few minutes on a single core. The median time required for model building with MODELLER-CSA and refinement was 6.2 and 1.1 hours, respectively, when 32 cores were used in parallel.
In this article, we report a new TBM method—GalaxyTBM—that builds reliable core structure from multiple templates and reconstructs unreliable regions by ab initio loop or terminus modeling in the fixed framework of the core structure. The current multiple-template strategy maximizes the positive effects of using multiple templates by selecting complementary multiple templates that do not contaminate the information from the best template significantly and by thorough optimization of possibly conflicting template restraints during model building. When model refinement by detection and re-building of unreliable loop or terminus regions is applied, additional improvement in model quality is observed. Several sound elements of the current strategy, such as template rescoring, multiple-template selection based on core-structure alignment, and multiple sequence alignment of core sequences could be easily incorporated into other TBM methods to enhance their performance.
The probability bins and the corresponding weight factors for rescoring were determined by maximizing the qualities of the top templates for the CASP8 TBM targets using a grid search in pre-set ranges of the parameters.
Multiple templates were then selected from the 20 top-ranked proteins as follows. First, the top 20 proteins were divided into high-rankers (those with the score S within 95% of the top ranker’s score) and low-rankers (the remaining ones). Second, those proteins that had dissimilar structures from a “background pool” of structures were removed. The background pool consisted of either the high-rankers or top 3 rankers, whichever was the greatest. The similarity of a protein structure to the background pool was measured by the mean TM-score from the pool structures, and the proteins that had lower similarity than the cut-off value, , were removed, where mpool and are the average and standard deviation of the similarity within the pool, respectively. The parameter α was set to 1 for the high-rankers and to the ratio of the protein’s S score to that of the top ranker for the low-rankers. When calculating TM-score between 2 protein structures, only the residues aligned to the target sequence by HHalign were considered, and the target sequence length was used as the reference length. Finally, proteins dissimilar from the top ranker, with TM-score < 0.5, were removed , where the sequence length of the top ranker was used as the reference length for TM-score calculation.
Multiple sequence alignment
Alignment between the target sequence and the template sequences was generated using PROMALS3D , one of the best multiple sequence alignment (MSA) tools available. PSI-BLAST  2.1.14 was used with default parameters (number of iterations = 3, e-value cut = 0.001) for sequence profile generation. TM-align  and DaliLite  were used as structure-alignment tools to provide the 3D structure information required for PROMALS3D. Default values were used for all the other parameters of PROMALS3D. Less meaningful terminus regions were temporarily neglected in the initial MSA and attached afterwards. The less meaningful regions were defined as the termini of query sequence not aligned to any templates by the global alignment using HHsearch. We did not take those termini into consideration at this point because we assumed that they could be modeled reliably in the later ab initio refinement stage. By neglecting those regions, the alignment effort was focused on the more reliable core region, increasing the possibility of generating a more reliable model structure for the core.
Model construction and optimization
Using the template structures and the MSA as input, a template-based model was constructed with the MODELLER-CSA  module, newly implemented in the GALAXY program package [11, 14–16]. MODELLER-CSA is a template-based model-building procedure that carries out global optimization of the MODELLER restraint function  using conformational space annealing (CSA) [28–30]. In the new implementation in GALAXY, the MODELLER restraints are interpreted in the source code level and local minimization in the CSA procedure is performed by a quasi-Newton minimizer . Both of these aspects are more advanced than the original implementation by Joo and coworkers . A typical run of model building generated 100 structures that maximally satisfy the restraints. Among the 100 structures, the structure nearest to the largest cluster center was selected as a representative model.
ULR detection and reconstruction
In the final stage, the single best template-based model structure was extensively refined by GalaxyREFINE , a high-resolution refinement method that employs advanced loop and terminus modeling algorithms [14, 32, 33]. Details of the refinement method can be found in Ref. 11, and here we describe it only briefly. ULRs were detected by a model-consensus quality assessment method . The conformational space of ULR was then searched using a global optimization procedure that combines triaxial loop closure [32, 33] and CSA on a newly introduced free energy surface composed of molecular mechanics force field , atomic-resolution statistical potential terms [35, 36], and additional supporting terms. Information from templates was not used for scoring in the refinement procedure. All the energy components and the sampling algorithms were implemented in the GALAXY program.
In the current application of GalaxyREFINE to the model refinement in GalaxyTBM, a few modifications were made to enhance the computational efficiency over that of the original version used in CASP9. First, ULRs detected by the model consensus method  were subject to a filtering scheme that eliminates ULRs with less than 6 or more than 20 residues. Out of the remaining ULRs, up to 3 ULRs with lowest reliability (largest fluctuations among generated models) were subjected to actual reconstruction. Another change was that multiple ULRs were refined simultaneously in a single optimization procedure, while separate optimization for each ULR was performed and the results were merged into a single structure in CASP9. Finally, the initial loop structures for CSA were generated by a slightly different method from that used in CASP9. While all 30 starting loop structures were generated de novo by FALC  in CASP9, 5 loops were taken from initial template-based models and 25 loops were generated by FALC in the current implementation. This modification indirectly accounts for template information, which can be helpful when regions with reliable templates are assigned as ULRs.
P-values were obtained from paired two-tailed Student’s t-test.
This work was supported by the KOSEF/MEST Grant No. 2011–0012456 and the Center for Marine Natural Products and Drug Discovery (CMDD), one of the MarineBio21 programs funded by the Ministry of Land, Transport, and Maritime Affairs.
- Zhang Y: Progress and challenges in protein structure prediction. Curr Opin Struct Biol 2008, 18: 342–348. 10.1016/j.sbi.2008.02.004PubMed CentralView ArticlePubMedGoogle Scholar
- Marti-Renom MA, Stuart AC, Fiser A, Sanchez R, Melo F, Sali A: Comparative protein structure modeling of genes and genomes. Annu Rev Biophys Biomol Struct 2000, 29: 291–325. 10.1146/annurev.biophys.29.1.291View ArticlePubMedGoogle Scholar
- Cheng J: A multiple-template combination algorithm for protein comparative modeling. BMC Struct Biol 2008, 8: 18. 10.1186/1472-6807-8-18PubMed CentralView ArticlePubMedGoogle Scholar
- Larsson P, Wallner B, Lindahl E, Elofsson A: Using multiple templates to improve quality of homology models in automated homology modeling. Protein Sci 2008, 17(6):990–1002. 10.1110/ps.073344908PubMed CentralView ArticlePubMedGoogle Scholar
- Venclovas C, Margelevicius M: Comparative modeling in CASP6 using consensus approach to template selection, sequence-structure alignment, and structure assessment. Proteins 2005, S7: 99–105.View ArticleGoogle Scholar
- Fernandez-Fuentes N, Rai BK, Madrid-Aliste CJ, Eduardo Fajardo J, Fiser A: Comparative protein structure modeling by combining multiple templates and optimizing sequence-to-structure alignments. Bioinformatics 2007, 19: 2558–2565.View ArticleGoogle Scholar
- Peng J, Xu J: Boosting protein threading accuracy. Research in Computational Molecular Biology 2009, 31–45.View ArticleGoogle Scholar
- Yang Y, Faraggi E, Zhao H, Zhou Y: Improving protein fold recognition and template-based modeling by employing probabilistic-based matching between predicted one-dimensional structural properties of the query and corresponding native properties of templates. Bioinformatics 2011, 27: 2076–2082. 10.1093/bioinformatics/btr350PubMed CentralView ArticlePubMedGoogle Scholar
- Hildebrand A, Remmert M, Biegert A, Soding J: Fast and accurate automatic structure prediction with HHpred. Proteins 2009, 77(S9):128–132. 10.1002/prot.22499View ArticlePubMedGoogle Scholar
- Lance BK, Deane CM, Wood GR: Exploring the potential of template-based modelling. Bioinformatics 2010, 26(15):1849–1856. 10.1093/bioinformatics/btq294View ArticlePubMedGoogle Scholar
- Park H, Seok C: Refinement of Unreliable local regions in template-based protein models. Proteins 2012, 80: 1974–1986.PubMedGoogle Scholar
- Cheng J, Eickholt J, Wang Z, Deng X: Recursive protein modeling: a divide and conquer strategy for protein structure prediction and its case study in CASP9. J Bioinform Comput Biol 2012, 10: 1242003. 10.1142/S0219720012420036PubMed CentralView ArticlePubMedGoogle Scholar
- Zhang Y: I-TASSER server for protein 3D structure prediction. BMC Bioinformatics 2008, 9: 40–47. 10.1186/1471-2105-9-40PubMed CentralView ArticlePubMedGoogle Scholar
- Lee J, Lee D, Park H, Coutsias EA, Seok C: Protein loop modeling by using fragment assembly and analytical loop closure. Proteins 2010, 78: 3428–3436. 10.1002/prot.22849PubMed CentralView ArticlePubMedGoogle Scholar
- Shin W, Heo L, Lee J, Ko J, Seok C, Lee J: LigDockCSA: protein-ligand docking using conformational space annealing. J Comput Chem 2011, 32: 3226–3232. 10.1002/jcc.21905View ArticlePubMedGoogle Scholar
- Park H, Ko J, Joo K, Lee J, Seok C, Lee J: Refinement of protein termini in template-based modeling using conformational space annealing. Proteins 2011, 79: 2725–2734. 10.1002/prot.23101View ArticlePubMedGoogle Scholar
- Soding J: Protein homology detection by HMM-HMM comparison. Bioinformatics 2005, 21: 951–960. 10.1093/bioinformatics/bti125View ArticlePubMedGoogle Scholar
- Pei J, Kim BH, Grishin N: PROMALS3D: a tool for multiple protein sequence and structure alignments. Nucleic Acids Res 2008, 36: 2295–2300. 10.1093/nar/gkn072PubMed CentralView ArticlePubMedGoogle Scholar
- Mariani V, Kiefer F, Schmidt T, Haas J, Schwede T: Assessment of template based protein structure predictions in CASP9. Proteins 2011, 79: 37–58. 10.1002/prot.23177View ArticlePubMedGoogle Scholar
- Zhang Y, Skolnick J: TM-align: A protein structure alignment algorithm based on TM-score. Nucleic Acids Res 2005, 3: 2302–2309.View ArticleGoogle Scholar
- Sali A, Blundell TL: Comparative protein modelling by satisfaction of spatial restraints. J Mol Biol 1993, 234(3):779–815. 10.1006/jmbi.1993.1626View ArticlePubMedGoogle Scholar
- Zemla A: LGA: a method for finding 3D similarities in protein structures. Nucleic Acids Res 2003, 31: 3370–3374. 10.1093/nar/gkg571PubMed CentralView ArticlePubMedGoogle Scholar
- Zemla A, Venclovas C, Moult J, Fidelis K: Processing and analysis of CASP3 protein structure predictions. Proteins 1999, S3: 22–29.View ArticleGoogle Scholar
- Joo K, Lee J, Lee K, Kim BG, Lee J: All-atom chain-building by optimizing MODELLER energy function using conformational space annealing. Proteins 2008, 75: 1010–1023.View ArticleGoogle Scholar
- Xu J, Zhang Y: How significant is a protein structure similarity with TM-score = 0.5? Bioinformatics 2010, 26: 889–895. 10.1093/bioinformatics/btq066PubMed CentralView ArticlePubMedGoogle Scholar
- Altchul SF, Madden TL, Scharffer AA, Zhang J, Zhang Z, Miller W, Lipman DJ: Gapped BLAST and PSI-BLAST: a new generation of protein database search programs. Nucleic Acids Res 1997, 25: 3389–3402. 10.1093/nar/25.17.3389View ArticleGoogle Scholar
- Holm L, Park J: DaliLite workbench for protein structure comparison. Bioinformatics 2000, 16: 566–567. 10.1093/bioinformatics/16.6.566View ArticlePubMedGoogle Scholar
- Lee J, Liwo A, Scheraga HA: Energy-based de novo protein folding by conformational space annealing and an off-lattice united-residue force field: application to the 10–55 fragment of staphylococcal protein a and to apo calbindin d9k. Proc Natl Aca Sci USA 1999, 2025–2030.Google Scholar
- Lee J, Scheraga HA, Rackovsky S: New optimization method for conformational energy calculations on polypeptides: Conformational space annealing. J Comput Chem 1997, 18: 1222–1232. 10.1002/(SICI)1096-987X(19970715)18:9<1222::AID-JCC10>3.0.CO;2-7View ArticleGoogle Scholar
- Lee J, Scheraga HA, Rackovsky S: Conformational analysis of the 20-residue membrane-bound portion of melittin by conformational space annealing. J Comput Chem 1998, 18: 1222–1232.View ArticleGoogle Scholar
- Liu D, Nocedal J: On the limited memory BFGS method for large scale optimization. Math Programming B 1989, 45: 503–528. 10.1007/BF01589116View ArticleGoogle Scholar
- Coutsias EA, Seok C, Jacobson MP, Dill K: A kinematic view of loop closure. J Comput Chem 2004, 25: 510–528. 10.1002/jcc.10416View ArticlePubMedGoogle Scholar
- Coutsias EA, Seok C, Wester MJ, Dill K: Resultants and loop closure. Int J Quantum Chem 2006, 106: 176–189. 10.1002/qua.20751View ArticleGoogle Scholar
- MacKerell AD Jr, Bashford D, Bellott M, Dunbrack RL Jr, Evanseck JD, Field MJ, Fischer S, Gao J, Guo H, Ha S, Joseph-McCarthy D, Kuchnir L, Kuczera K, Lau FTK, Mattos C, Michnick S, Ngo T, Nguyen DT, Prodhom B, Reiher WE III, Roux B, Schienkrich M, Smith JC, Stote R, Straub J, Watanabe M, Wiorkiewicz-Kuczera J, Yin D, Karplus M: All-atom empirical potential for molecular modeling and dynamics studies of proteins. J Phys Chem B 2002, 102: 3586–3616.View ArticleGoogle 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 Sci 2002, 11: 2714–2726.PubMed CentralView ArticlePubMedGoogle Scholar
- Yang Y, Zhou Y: Ab initio folding of terminal segments with secondary structures reveals the fine difference between two closely-related all-atom statistical energy functions. Protein Sci 2008, 17: 1212–1219. 10.1110/ps.033480.107PubMed CentralView ArticlePubMedGoogle Scholar
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.