Template-based modeling that employs various meta-threading techniques is currently the most accurate, and consequently the most commonly used, approach for protein structure prediction. Despite the evident progress in this field, accurate structure models cannot be constructed for a significant fraction of gene products, thus the development of new algorithms is required. Here, we describe the development, optimization and large-scale benchmarking of eThread, a highly accurate meta-threading procedure for the identification of structural templates and the construction of corresponding target-to-template alignments. eThread integrates ten state-of-the-art threading/fold recognition algorithms in a local environment and extensively uses various machine learning techniques to carry out fully automated template-based protein structure modeling. Tertiary structure prediction employs two protocols based on widely used modeling algorithms: Modeller and TASSER-Lite. As a part of eThread, we also developed eContact, which is a Bayesian classifier for the prediction of inter-residue contacts and eRank, which effectively ranks generated multiple protein models and provides reliable confidence estimates as structure quality assessment. Excluding closely related templates from the modeling process, eThread generates models, which are correct at the fold level, for >80% of the targets; 40–50% of the constructed models are of a very high quality, which would be considered accurate at the family level. Furthermore, in large-scale benchmarking, we compare the performance of eThread to several alternative methods commonly used in protein structure prediction. Finally, we estimate the upper bound for this type of approach and discuss the directions towards further improvements.
Citation: Brylinski M, Lingam D (2012) eThread: A Highly Optimized Machine Learning-Based Approach to Meta-Threading and the Modeling of Protein Tertiary Structures. PLoS ONE 7(11): e50200. doi:10.1371/journal.pone.0050200
Editor: Lukasz Kurgan, University of Alberta, Canada
Received: August 4, 2012; Accepted: October 22, 2012; Published: November 21, 2012
Copyright: © 2012 Brylinski, Lingam. 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 study was supported by the Louisiana Board of Regents through the Board of Regents Support Fund [contract LEQSF(2012-15)-RD-A-05]. Portions of this research were conducted with high performance computational resources provided by Louisiana State University (HPC@LSU, http://www.hpc.lsu.edu) and the Louisiana Optical Network Institute (LONI, http://www.loni.org). The funders 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.
With the continuing advances in genome sequencing , there has been a rapid accumulation of protein sequences, whose structures are yet to be annotated. As of October 2012, there are >1.7×107 unique protein sequences from 17,994 organisms in the Reference Sequence database . However, due to low-sequence identity to already annotated proteins, the molecular functions of many of these gene products remain unknown. Using standard homology-based tools poses a significant risk associated with the “overprediction” of molecular function and, as an inevitable consequence, typically results in high levels of misannotation . On that account, more accurate and confident function annotation tools are needed; here structure-based approaches show a considerable promise . Early methods for function inference from protein structure were very sensitive to the quality of the target structures and typically required these solved experimentally by X-ray crystallography or NMR. More recent approaches are generally devoid of these limitations and can routinely annotate low-to-moderate quality protein models , , , . Consequently, protein structure modeling plays an important role in Functional Genomics by providing structural information on gene products that is subsequently utilized by powerful structure-based approaches to protein function inference , .
Currently, the most accurate and the most widely used methods for protein structure prediction build on homology, i.e. they use information educed from related proteins. As demonstrated in the recent community-wide Critical Assessment of Protein Structure Prediction (CASP) experiment, the top performing groups in tertiary structure prediction category used various template-based methods . One of the best algorithms in the field, I-TASSER, builds three-dimensional models from multiple-threading alignments constructed by LOMETS  using iterative assembly/refinement simulations ; this is followed by function prediction by matching the models to proteins with known functions . Another development from this prolific group is QUARK, a method for protein structure assembly using continuous template fragments . QUARK first identifies small structural fragments by gapless threading against the Protein Data Bank (PDB)  and then ranks them using a composite scoring function, which consists of sequence and structure profiles, predicted secondary structure and backbone torsion angles. For each position in the target, the top-scored fragments are used to assemble a 3D model by Replica Exchange Monte Carlo simulations. Recent improvements of template selection methods include the development of HHblits, a new iterative HMM-HMM sequence search algorithm . HHblits was demonstrated to have 50–100% higher sensitivity than PSI-BLAST  and to produce multiple alignments of much higher quality. Furthermore, advances in the quality assessment protocol result in a significant gain in the overall performance of IntFOLD-TS, which first generates a large number of alternate models using in-house versions of several different alignment methods and then ranks them in terms of the estimated global quality . Importantly, highly accurate predictions of local errors, provided in the resulting models, make this method useful for guiding future experimental work. Improved prediction of secondary structure, backbone torsion angles and solvent accessible surface area significantly increases the accuracy of SPARKS-X, which is one of the best single-method fold recognition techniques . Finally, RaptorX uses a novel statistical learning model and a multiple-template threading component to provide better measure of the compatibility between the target sequence and the template structures . Indeed, the constructed alignments are much more accurate than those built by its predecessor, RAPTOR. These and many other successful examples show that there is an encouraging progress in this field, which certainly will have impact on many areas of modern molecular and cell biology.
Notwithstanding the success of single-threading approaches, meta-threading techniques are the ones that make headway in protein structure prediction. These methods identify template structures and construct target-to-template alignments by considering outputs from a variety of individual threading algorithms. Typically, the combined predictions have a higher chance to be accurate than those produced by a single method. Recent CASP experiments demonstrated that models generated from predictions by meta-threading servers are more accurate than the best individual server alone , . Moreover, an important additional advantage of meta-predictors is the improved estimation of the reliability of predictions. An example of such a successful meta-server is LOMETS, which currently uses ten threading algorithms to generate initial structural models and constraints for the prediction of protein tertiary structures . Models in LOMETS are selected from individual programs purely based on consensus, i.e. the structure similarity of the considered model with other threading alignments. The consensus predictions provided by LOMETS were shown to be more accurate than those generated by individual component methods. Another example is Pcons, a neural-network–based consensus predictor that improves fold recognition by selecting the best model out of those produced by six prediction servers . Pcons translates the confidence scores reported by each server into uniformly scaled values corresponding to the expected accuracy of each model. The translated scores as well as the similarity between models produced by different servers are used in the final selection. According to benchmarks carried out for two unrelated sets of newly solved proteins, Pcons outperforms any single server.
In this communication, we describe eThread, a highly accurate meta-threading procedure to identify templates for the template-based modeling of protein structures. This new method uses ten state-of-the-art threading algorithms and machine learning designed specifically for the optimal selection of structure templates. In large-scale benchmarks, we demonstrate that the performance of eThread in the identification of structurally related templates is notably higher than any of the individual single-threading algorithms. Template-based protein structure modeling requires not only a set of structure templates but also the corresponding target-to-template alignments and/or predicted inter-residue contacts. Therefore, as a part of eThread software, we developed a new machine learning procedure to combine alignments reported by individual meta-threading algorithms into a set of consensus alignments. We also developed eContact, a Bayesian classifier with an optimized Gaussian kernel for the prediction of inter-residue contacts. Optimized sets of templates and the corresponding alignments as well as predicted long-range contacts are integrated into structure assembly protocols for the construction of full-length models of the target proteins. Two separate procedures have been devised based on widely used modeling algorithms: Modeller  and TASSER-Lite . In addition, we designed eRank, which effectively ranks generated multiple protein models and provides reliable confidence estimates for structure quality assessment. To demonstrate the utility of this approach, modeling protocols were optimized and carefully benchmarked on a large and representative dataset of protein structures and compared to the performance of several alternative methods commonly used in protein structure prediction. Finally, we estimate the upper bound for this type of approach and discuss the directions towards further improvements. eThread webserver as well as benchmarking datasets and results are freely available to the academic community at http://www.brylinski.org/ethread.
Materials and Methods
A flowchart for the eThread algorithm is shown in Figure 1. For a given amino acid sequence, the method starts by applying meta-threading to search for structurally similar templates in two libraries, which consist of full protein chains as well as individual domains. The inclusion of individual domains is a commonly used practice in threading to improve the recognition of those templates that may only partially cover a multiple-domain target . In addition, if a full chain template is found, it also provides the information on the mutual orientation of domains. The identified templates are subsequently filtered by eThread and the corresponding target-to-template alignments are constructed. Next, two structure modeling protocols are used to build the three-dimensional models of the target: Modeller , which employs template pre-clustering by MaxCluster, and TASSER-Lite , which additionally incorporates inter-residue contacts predicted by eContact. In both cases, the resulting models are ranked by eRank, assigned confidence estimates, and refined using molecular mechanics. Below is a detailed description of the benchmarking dataset as well as the individual modeling stages.
Benchmarking dataset was compiled from all PDB entries as of Jan 2012. Using PISCES  to remove redundancy at the 40% sequence identity resulted in 11,468 representative protein chains 50–600 residues in length. Furthermore, we excluded 2,596 proteins, for which no structurally related proteins can be detected using any of the individual threading component methods. The final dataset consists of 8,872 non-redundant and representative protein targets.
Two threading libraries are used in this study: chain and domain. Chain library comprises aforementioned 11,468 protein chains selected from the PDB by PISCES . Domain library was compiled by PISCES using the Structural Classification of Proteins (SCOP) database . Similarly to the chain library, the redundancy was removed at 40% pairwise sequence identity. This library contains 10,013 representative protein domains 50–600 residues in length, for which the atomic coordinates were obtained from the ASTRAL database .
Threading Component Methods and Template Selection
eThread is a meta-threading procedure, which integrates ten state-of-the-art protein threading/fold recognition algorithms: CSI-BLAST , COMPASS , HHpred , HMMER , pfTools , pGenThreader , SAM-T2K , SP3 , SPARKS2  and Threader . Each individual threading/fold recognition algorithm assesses structures present in the template library using some scoring system, e.g. SP3, SPARKS2 and Threader assign Z-scores using the entire template library as a background, COMPASS, CSI-BLAST, HMMER and SAM-T2K employ scoring systems based on analytically estimated E-values, and HHpred uses calibrated probabilities for true relationships between proteins. For the template selection, we constructed a machine learning model based on feature vectors composed of individual threading scores. The machine learning employs Support Vector Machines for classification problems (SVC)  to assess whether a particular template is structurally related to the target with a TM-score  of ≥0.4. The accuracy of template selection is assessed using 2-fold cross validation excluding those templates, whose sequence identity to target is >40%. We note that this sequence identity cutoff is also applied in all subsequent modeling steps.
Consensus Target-to-template Alignments
As a part of eThread, we also developed a new machine learning procedure to combine alignments reported by individual meta-threading algorithms into a set of consensus alignments. Specifically, we built a Naïve Bayes classifier, which was trained on meta-threading data against reference structure alignments constructed by fr-TM-align . First, from individual alignments produced by the component methods, this model estimates the posterior probability of each pair of residues to be a part of the target-to-template structure alignment. Subsequently, the matrix of Bayesian probabilities is used as a scoring function in Needleman-Wunsch Dynamic Programming (DP)  to construct the consensus global alignments. Similarly to the template selection, the consensus alignment model is assessed using 2-fold cross validation.
Inter-residue Contact Prediction
Long-range contacts between residues are defined when a pair of their heavy atoms is within a distance of 4.5 Å and they are separated in the sequence by at least 4 other residues. Inter-residue contacts are predicted from consensus target-to-template alignments by eContact, a machine learning approach. For a pair of residues, we calculate a vector of four features: the fraction of templates that have residues in equivalent positions in contact with each other, the average confidence of these templates that have such contacting residues, and the average confidence of the corresponding target-to-template alignments; in addition, we also include a knowledge-based statistical pair potential . Based on these feature vectors, a SVC  model was constructed to estimate the probability of a given pair of residues to be in contact. The accuracy was assessed by 2-fold cross validation.
Tertiary Structure Modeling
To construct three-dimensional models of the target proteins, we employ two commonly used template-based modeling algorithms: Modeller  and TASSER-Lite . Both programs use threading alignments generated by eThread as input. In addition, TASSER-Lite also uses inter-residue contacts predicted by eContact. For Modeller, the set of templates identified by eThread is pre-clustered by MaxCluster (http://www.sbg.bio.ic.ac.uk/maxcluster/) using a TM-score clustering threshold of 0.4 and the models are subsequently constructed individually for each cluster. The side chains in the structures modeled by Modeller and TASSER-Lite are rebuilt from the Cα trace by Pulchra  and finally, all-atom structures are refined in the CHARMM22 force field  using the Jackal modeling package .
Model Ranking and Confidence Estimates
Both Modeller and TASSER typically generate multiple models for a given target. To rank the resulting models and to assign confidence estimates, we developed eRank that employs SVM-Rank, a version of Support Vector Machines designed specifically for ranking problems . eRank also estimates the TM-score to native using Support Vector Regression (SVR) . Both ranking and confidence estimate models use the following set of features: the confidence of alignments constructed by eThread (Alignment), the average alignment coverage (Coverage), DOPE score  (DOPE), dFire residue-level potential of mean force  (dFire), secondary structure match between the model and the PSIPRED  prediction (PSIPRED), burial score (Burial, see Text S1) and secondary structure preferences (SecStr, see Text S1). In addition, eRank/Modeller includes the fraction of templates assigned to a particular cluster by MaxCluster (Fraction) and the GA341 score  (GA341). eRank/TASSER-Lite also incorporates the average TM-score of a given model to templates identified by eThread (TM-score, roughly equivalent to Fraction for Modeller) as well as the following clustering coefficients reported by SPICKER , which is a part of the TASSER-Lite suite: cluster fraction (TASSER-Litefract), cluster density (TASSER-Litedens) and cluster mean energy (TASSER-Liteene). Both ranking ability and the accuracy of confidence estimates are assessed using 2-fold cross validation.
Other Approaches to Structure Modeling
We compare the accuracy of eThread models to those constructed by two alternative protocols. The first one is a naïve, single-template approach: For a given target sequence, we run 5 iterations of PSI-BLAST  to identify weakly homologous proteins and we select the top-ranked as the structure template. A three-dimensional model is then constructed by Nest  using the target-to-template alignment provided by PSI-BLAST. The resulting model is additionally subject to all-atom structure refinement using the Jackal modeling package . The second approach represents a single-threading, multiple-template algorithm; here, we use the original implementation of TASSER-Lite . For both PSI-BLAST/Nest and TASSER-Lite, we exclude from the modeling procedure all closely related templates with >40% sequence identity to the target in order to make the results comparable to those obtained by eThread-based modeling.
The ability of a threading algorithm to select those templates that are structurally similar to the target is critical for the subsequent construction of three-dimensional models. Here, we define a good template as the structure with a statistically significant TM-score  to native of ≥0.4. We note that a TM-score of 0.4 is an appropriate fold similarity assignment threshold; template structures above this value contain sufficient information to enable the full-length reconstruction of the target structure . TM-score is calculated by fr-TM-align  for both threading libraries used in this study. Trivial templates with more than 40% sequence identity to target are excluded from this as well as all subsequent analyses. Figure 2 shows ROC plots for eThread compared to the individual threading component methods. The accuracy does not depend on the library used (Figure 2A – chain, Figure 2B – domain); however, it varies across different algorithms. HHpred was found the most accurate single method with a true positive rate (TPR) of 0.49/0.50 at the expense of 0.05 false positive rate (FPR) for the chain/domain library. At the same FPR, the next accurate algorithms: SP3, COMPASS and SPARKS2 give a TPR of 0.47/0.43, 0.44/0.41 and 0.44/0.40, respectively. However, the effective combination of multiple algorithms considerably extends the coverage of target sequences by distantly related templates and increases the true positive rate; the corresponding TPR values for eThread are 0.60/0.57 (at 0.05 FPR). Thus eThread systematically detects more structure templates than any of the component methods. The probability values returned by machine learning also contribute to the overall modeling confidence.
Quality of Threading Alignments
Effective template selection is still not sufficient for practical applications, such as protein structure modeling. In addition, target-to-template alignments should also be accurate to build a correct model. In Figure 3, we assess the quality of threading alignments constructed by eThread as well as all component methods by Matthew’s correlation coefficient (MCC) against structure alignments by fr-TM-align. Again, HHpred, SP3 and SPARKS2 were found to be the most effective single-threading algorithms that build alignments with a MCC of ≥0.5 for the chain (domain) library for 69% (71%), 65% (68%) and 63% (66%) of the targets, respectively. The performance of eThread is slightly lower than that of HHpred for MCC>0.6; however, it still provides good quality alignments in the MCC range of 0.4–0.6 for additional 4–8% of the targets, on average.
Accuracy of Inter-residue Contacts
In addition to threading templates and target-to-template alignments, TASSER-Lite also incorporates predicted inter-residue contacts as an important component of its force field . Here, we developed eContact, a machine learning-based method for the prediction of long-range contacts. eContact uses threading alignments as well as a generic knowledge-based pair potential; its cross-validated performance on a representative dataset is shown in Figure 4. At least 75% of exact native contacts are recovered for 72% of the target proteins. To select the optimal cutoff value for contact prediction, we use MCC, which represents a balanced measure that can be used if the classes are of different sizes . The contact probability threshold of 0.35 maximizes MCC to 0.65 against the exact native contacts and yields 0.79 of true positive rate at the expense of only 0.14 false positives (Figure 4 inset). The accuracy further increases, when contacts within 1, 2 and 3 residues are also considered positives. Here, the fraction of targets with ≥75% of predicted native contacts is 84%, 88% and 91%, respectively. We note that TASSER-Lite, which employs low-resolution modeling, can effectively accommodate inter-residue contacts slightly mispredicted by a couple of residues.
The accuracy is evaluated for exact contacts as well as those within 1, 2 and 3 residues from the exact contact. Inset: ROC plot for the contact prediction; TPR – true positive rate, FPR – false positive rate. Star corresponds to a contact probability that maximizes MCC for eContact, gray triangle depicts the performance of SVMcon and the dotted line shows the TPR improvement of eContact over SVMcon for the FPR fixed at 0.047.
We also compare the performance of eContact to that of SVMcon, ranked as one of the top residue contact predictors in CASP7 . SVMcon employs machine learning and a set of features, which include sequence profiles, secondary structure, solvent accessibility and contact potentials . At a fixed FPR rate of 0.047, eContact and SVMcon yield TPR of 0.62 and 0.44, respectively (Figure 4 inset); thus eContact predicts 18% more contacts than SVMcon.
Both Modeller and TASSER-Lite typically build multiple models. For Modeller, we first pre-cluster the set of templates identified by eThread and then construct a structural model for each cluster. TASSER-Lite generates Monte Carlo trajectories, which are subsequently clustered by SPICKER and a structure closest to the cluster centroid is selected for each cluster. To select the best models, we developed eRank/Modeller and eRank/TASSER-Lite; both are machine learning approaches that use a variety of scoring functions. In Figure 5, we assess the ranking ability of eRank, i.e. in how many cases the best model is found amongst the top 5 ranks; we also compare the performance of eRank to the component scoring functions. We note that the “best model” may not be necessarily highly accurate; it is just better than the other models constructed. As shown in Figure 5A, eRank/Modeller correctly identifies the best model in 95% of the cases, which represents an improvement over the most effective individual scoring terms: DOPE (87%), dFire (86%), Fraction (81%), PSIPRED (79%) and Coverage (78%). eRank/TASSER-Lite ranks the best model as the 1st, 2nd and 3rd one in 41%, 33% and 17% of the cases, respectively (see Figure 5B). Again, this ranking accuracy is higher than TASSER-Litedens, TASSER-Litefract and TASSER-Liteene, which place the best model at rank 1 for 38%, 38% and 37% of the targets, respectively.
Structure models constructed by (A) Modeller and (B) TASSER-Lite are ranked and the corresponding accuracy is assessed by the fraction of targets for which the best models was found at a particular rank. Dashed black and dotted/dashed gray line depicts ranking accuracy by SELECTpro and APOLLO, respectively.
In Figure 5, eRank is also compared to SELECTpro  (dashed black) and APOLLO  (dotted/dashed gray), which are structure-based model selection methods. SELECTpro uses a sophisticated energy function that comprises physical, statistical and predicted structural scoring terms and was shown in large-scale benchmarks to be highly effective. APOLLO evaluates the absolute single and pair-wise global structure quality in terms of the GDT-score ; here, we use the single-model approach. eRank outperforms SELECTpro for the models constructed by Modeller (TASSER-Lite) by ∼10%; here, the best model is assigned rank 1 in 95% (41%) and 85% (32%) of the cases, respectively. The accuracy of APOLLO is slightly higher than SELECTpro; however, ∼8% worse than eRank: the best eThread/Modeller (eThread/TASSER-Lite) model is assigned rank 1 in 86% (33%) of the cases. We note that eRank was specifically tailored to structure modeling using eThread, whereas SELECTpro and APOLLO represent general quality assessment approaches, applicable to any set of protein models. The difference in performance between eRank/Modeller and eRank/TASSER-Lite comes from the way models are constructed and from the pairwise similarities between the top-ranked structures. Multiple TASSER-Lite models are often structurally similar to each other (as well as to the target), thus the ranking is more difficult. The pre-clustering procedure used in the model construction by Modeller typically results in a set of very different structures with a pairwise TM-score of <0.4; consequently, at most one model would be structurally similar to the target.
We use TM-score  to native as the main assessment metric for the accuracy of the top-ranked models. Note that the TM-score is a protein length independent measure of structural similarity with a statistical significance at ≥0.4. In addition, we assess the structure quality using several other well established measures: Cα-RMSD , Gaussian-weighted RMSD (wRMSD) , MaxSub  and GDT-score . Benchmarking results reported here were obtained for a non-redundant and representative subset of the PDB; therefore are easily comparable to other studies that use a similar setup. Moreover, these statistics provide reliable estimates of the expected accuracy in large-scale applications, e.g. genome-wide protein structure modeling projects .
For both Modeller and TASSER-Lite, we also evaluate the models constructed using 3 different protocols to ascertain, where the future improvements are most likely to increase the overall accuracy of structure modeling. First, we assess the complete eThread procedure, i.e. template identification, alignment construction and model assembly/refinement. Next, to evaluate the quality of target-to-template alignments, we include only these templates that are structurally related to the target with a TM-score of ≥0.4. Finally, we estimate the upper bound for the modeling accuracy using structurally similar templates only and the corresponding structure alignments constructed by fr-TM-align. The results for Modeller and TASSER-Lite are shown in Figure 6 as a fraction of targets whose structures are modeled to a given accuracy. Focusing on a high modeling accuracy at a TM-score of ≥0.7, the upper bound for the modeling protocols using Modeller and TASSER-Lite is 78% and 75%, respectively. The accuracy of modeling using eThread alignments instead of these constructed by fr-TM-align (eThread/good templates in Figure 6) decreases to 54% (by 24%) and to 55% (by 20%) for Modeller and TASSER-Lite, respectively. It shows that TASSER-Lite better accommodates alignment errors than Modeller. When the complete eThread procedure is used, protein models with a TM-score of ≥0.7 are constructed by Modeller and TASSER-Lite for 49% and 39% of the targets, respectively. It demonstrates, that Modeller builds more highly accurate models; this is also shown in Table 1, which assesses the structure quality using several other measures. For example, the average MaxSub (GDT-score) for eThread/Modeller and eThread/TASSER-Lite is 0.55 (0.59) and 0.46 (0.50), respectively. However, Modeller provides slightly lower coverage of a dataset by models whose TM-score to native is ≥0.4 (still statistically significant) than TASSER-Lite: 85% and 88%, respectively (see Figure 6). Nevertheless, using eThread identified templates and alignments and model ranking by eRank, both structure modeling algorithms build correct (and often very high quality) models for a significant fraction of the benchmark proteins.
Three sets of (A) Modeller and (B) TASSER-Lite models are constructed using: structure alignments, eThread alignments for structurally related (“good”) templates only as well as all alignments generated by eThread. Gray-bricked area points up a room for further improvement using structure refinement. Dashed and dotted line corresponds to the accuracy of PSI-BLAST/Nest and the original TASSER-Lite, respectively.
This modeling accuracy is also higher than that obtained using a simple single-template approach, see Figure 6 and Table 1. For 77% of the target proteins, PSI-BLAST/Nest constructs models whose TM-score to native is ≥0.4. This is 8% and 11% less than using eThread/Modeller and eThread/TASSER-Lite, respectively. When compared to a single-threading, multiple-template approach, the most notable improvement is for protein models with a TM-score to native of ≥0.7. Here, the original TASSER-Lite generates models with such accuracy for 37% of the targets, which is 2% less than using eThread/TASSER-Lite; however, for 12% more target proteins high quality models are constructed using eThread/Modeller. This justifies the computationally more expensive multiple-template modeling using meta-threading and eThread.
Stereochemical Quality of Models
In addition to the global accuracy of protein models, we also assess their local stereochemical quality as reported by PROCHECK . Table 2 shows that the stereochemical quality of eThread models, particularly those constructed by eThread/Modeller, is quite high and very close to crystal structures. For example, only 5% less residues are assigned to the most favored regions on the Ramachandran map for the top-ranked models. Top-ranked eThread/TASSER-Lite models are ∼15% worse than these built by eThread/Modeller, suggesting that the former may require more rigorous local structure refinement. Furthermore, in both cases, the top-ranked models typically have higher stereochemical quality than those at lower ranks. Finally, both procedures, eThread/Modeller and eThread/TASSER-Lite, systematically produce models whose quality is notably higher than that obtained by a single-template approach, PSI-BLAST/Nest as well as the standard version of TASSER-Lite, see Table 2.
Model Quality Assessment
A modern structure modeling protocol also requires a reliable system to estimate the modeling confidence, which is often called quality assessment. eRank uses machine learning models appropriate for regression problems to provide this functionality. For a given model, the confidence corresponds to the estimated TM-score to native. Figure 7 shows correlation plots for the top five models constructed by eThread/Modeller. The Pearson correlation coefficient (CC) is used to measure the strength of a linear dependence between the predicted and real TM-score values. CC of 0.89 produced by eRank/Modeller (Figure 7A) is much higher than the individual scoring functions, e.g. Coverage (0.68), PSIPRED (0.63), dFire (0.60) or DOPE (0.56). eRank/TASSER-Lite also provides very reliable confidence estimates with a CC of 0.81 (Figure 8), despite the higher density of good models with a TM-score of ≥0.4. Here, the most accurate individual scoring functions, TASSER-Liteene, TM-score and DOPE are notably less accurate with the CC of 0.55, 0.49 and 0.49, respectively.
Three plots on the left show the correlation between the real TM-score of models built by eThread/Modeller and the TM-score estimated by (A) eRank, (B) SELECTpro and (C) APOLLO. For APOLLO, GDT-score is used instead of TM-score. Individual scoring components of eRank are shown on the right: (D) Fraction, (E) eThread, (F) DOPE, (G) GA341, (H) dFire, (I) PSIPRED, (J) SecStr, (K) Burial, (L) Alignment, and (M) Coverage. In A, B and D–M, dotted lines delineate the TM-score statistical significance threshold.
In both cases, the CC between predicted and real TM-score values for eRank is significantly higher than that obtained by an alternate model quality assessment method, SELECTpro , which produces the CC of 0.42 (Figure 7B) and 0.06 (Figure 8B) for eThread/Modeller and eThread/TASSER-Lite models, respectively. A common feature of structure-based methods, such as SELECTpro, DOPE or GA341 is that these algorithms typically recognize good models, but also assign high scores to non-native conformations, which are of acceptable stereochemical quality, e.g. Figure7B, 7E and 7F. In addition to SELECTpro, we also compare eRank to APOLLO  using a single-structure mode. Here, we switch to GDT-score, which is the default scoring function used by this algorithm. We note that the real GDT-scores calculated for eThread/Modeller and eThread/TASSER-Lite models correlate very well with the corresponding TM-scores (CC of 0.93 and 0.90, respectively). APOLLO builds more accurate estimates of the global structure quality than SELECTpro and all individual scoring functions. For models constructed by eThread/Modeller and eThread/TASSER-Lite, the CC is 0.77 and 0.65, respectively (see Figures 7C and 8C). Nonetheless, eRank still gives 12–16% higher correlation than APOLLO; thus the scoring function implemented in eRank clearly provides a robust system for the a priori estimate of model divergence from the native conformation.
Three plots on the left show the correlation between the real TM-score of models built by eThread/TASSER-Lite and the TM-score is estimated by (A) eRank, (B) SELECTpro and (C) APOLLO. For APOLLO, GDT-score is used instead of TM-score. Individual scoring components of eRank are shown on the right: (D) TM-score, (E) eThread, (F) Alignment, (G) Coverage, (H) DOPE, (I) dFire, (J) PSIPRED, (K) SecStr, (L) Burial, (M) TASSER-Litefract, (N) TASSER-Litedens, and (O) TASSER-Liteene. In A, B and D–O, dotted lines delineate the TM-score statistical significance threshold.
eThread Webserver and Datasets
eThread webserver, datasets and modeling results are available for non-commercial users under the terms of GNU General Public License at http://www.brylinski.org/ethread. The webserver allows users to submit amino acid sequences 50–600 residues in length, select the modeling protocol (either eThread/Modeller or eThread/TASSER-Lite) and download the results as well as visualize them directly on the website using OpenAstexViewer . The webserver was designed to be user-friendly and accessible using a Java-enabled web browser and any operating system.
Template-based modeling is currently the most accurate, and consequently the most commonly used, approach for protein structure prediction. The best methods in this area frequently employ meta-threading to identify template structures in available databases, such as PDB  and to construct target-to-template alignments. A popular technique used in the development of meta-threading approaches is a gateway approach, which queries several publicly available servers, collects the results and generates consensus predictions , , . For example, a neural-network approach that combines predictions from six webservers was demonstrated to increase the accuracy of fold recognition by 8–10% . Nevertheless, it was pointed out that the gateway approach may result in unexpected delays and possibly inconsistent results as a consequence of shutting down remote machines, frequent updates and modifications of algorithms used remotely . Thus, a meta-approach set up and maintained in a local environment appears as the most steady, robust and desirable solution.
In this work, we describe the development, optimization and large-scale benchmarking of eThread, a machine learning-based method, which integrates ten state-of-the-art threading/fold recognition algorithms in a local environment to carry out fully automated template-based protein structure modeling. Excluding closely related templates from the modeling process, we evaluate the performance of eThread in template identification, the construction of threading alignments and inter-residue contact prediction. We demonstrate that eThread generates high-quality structural data that can be effectively used to build reliable protein models using available structure assembly algorithms. eThread extensively uses various machine learning techniques to make highly accurate predictions. It has been demonstrated that statistical machine learning effectively utilizes a set of features extracted using general-purpose alignment tools for template ranking; here, sequence profile-profile and profile-structural-profile scores are the most informative , . A model based on Support Vector Machines was also successfully applied to estimate the significance of individual target-to-template alignments with a notable improvement over more standard measures such as Z-score or E-value . In our benchmarks, eThread detects significantly more templates than any single-threading algorithm while maintaining a low false-positive rate.
Next to template identification, the construction of corresponding target-to-template alignments is critical to produce high-quality protein models. Alignments generated by different methods can confidently cover different regions of a target sequence, thus the consensus alignment may result in a significant global improvement. Better-aligned parts recognized in a set of alignments generated by different methods can be combined into a unique solution, which is typically more accurate than any of the individual alignments , . Here, we developed a machine learning variant of this approach, which applies a Bayesian Classifier to meta-threading alignments to construct a probability-based scoring matrix, which is subsequently used in a traditional Needleman-Wunsch DP. The accuracy of this algorithm is comparable to the best individual alignment method for easy targets, but outperforms other methods in more difficult cases.
Furthermore, we developed eContact, a new machine learning-based method for inter-residue contact prediction that takes advantage of accurately identified templates, good quality target-to-template alignments and a knowledge-based statistical pair potential to recover native contacts at a very low false positive rate. In contact prediction, applying non-linear models, such as support vector machines, frequently outperforms many of the simple majority voting methods , . In addition to protein structure modeling, the predicted inter-residue contacts are also useful for the estimation of protein folding rates .
To assembly three-dimensional models of the target proteins using eThread templates and alignments, we tested two popular structure modeling algorithms: Modeller  and TASSER-Lite . Both programs perform comparably well and generate models, which are correct at the fold level, for >80% of the targets. However, significantly less of the constructed models (40–50%) are of a very high quality, which would be considered accurate at the family level . Here, the upper bound estimated using structure alignments is ∼75%, which suggest that further advances in threading methodologies could bring about 25% improvement in low-homology template-based modeling. Generating near-experimental quality structural models using “twilight zone” templates  would therefore require different modeling techniques, such as all-atom refinement , , . A gray-bricked area in Figure 6 points up a substantial room for potential improvement using structure refinement, which increases with the requirement of protein models to be closer to experimental structures. In this study, we employ a very simple procedure for all-atom refinement using molecular mechanics, which mostly optimizes side chain geometries and removes atom clashes. Using more advanced refinement could yield additional improvement in model quality, particularly in the high TM-score regime.
Many state-of-the-art protein structure prediction algorithms often generate a set of possible models for a given target. This is particularly common in low-homology multiple-template modeling. Thus, there is a need to select the most native-like conformation from a pool of constructed models. To address this issue, we developed eRank, which employs support vector machines for ranking problems to provide a very robust approach to model ranking. In addition, eRank also produces reliable confidence estimates, which correlate well with the actual model quality. This is particularly important for the use of modeled structures in structure-based function annotation. For example, in ligand and macromolecular docking, the selection of modeling protocol strongly depends on the quality of the target protein structures. While all-atom docking is applicable to high-quality receptor structures , , using low-to-moderate quality protein models often requires different algorithms, such as low-resolution modeling , , to provide confident annotations.
We present a suite of programs: eThread, eContact and eRank, which build on meta-threading and conduct fully automated template-based protein structure modeling. This meta-approach extensively uses machine learning techniques to generate good quality protein models even in the presence of only distantly homologous template structures and offers a reliable system for confidence estimates. Comparative benchmarks show that it outperforms other methods for inter-residue contact prediction, template-based structure modeling as well as model selection and quality assessment. eThread is freely available to the academic community through a user-friendly webserver at http://www.brylinski.org/ethread.
Calculation of the burial score and secondary structure preferences for model ranking and confidence estimates.
Conceived and designed the experiments: MB. Performed the experiments: MB. Analyzed the data: MB. Contributed reagents/materials/analysis tools: MB DL. Wrote the paper: MB.
- 1. Zhao J, Grant SF (2011) Advances in whole genome sequencing technology. Curr Pharm Biotechnol 12: 293–305.
- 2. Pruitt KD, Tatusova T, Klimke W, Maglott DR (2009) NCBI Reference Sequences: current status, policy and new initiatives. Nucleic Acids Res 37: D32–36.
- 3. Schnoes AM, Brown SD, Dodevski I, Babbitt PC (2009) Annotation error in public databases: misannotation of molecular function in enzyme superfamilies. PLoS Comput Biol 5: e1000605.
- 4. Skolnick J, Brylinski M (2009) FINDSITE: a combined evolution/structure-based approach to protein function prediction. Brief Bioinform 10: 378–391.
- 5. Roy A, Kucukural A, Zhang Y (2010) I-TASSER: a unified platform for automated protein structure and function prediction. Nat Protoc 5: 725–738.
- 6. Brylinski M, Skolnick J (2008) A threading-based method (FINDSITE) for ligand-binding site prediction and functional annotation. Proc Natl Acad Sci U S A 105: 129–134.
- 7. Oh M, Joo K, Lee J (2009) Protein-binding site prediction based on three-dimensional protein modeling. Proteins 77 Suppl 9152–156.
- 8. Wass MN, Sternberg MJ (2009) Prediction of ligand binding sites using homologous structures and conservation at CASP8. Proteins 77 Suppl 9147–151.
- 9. Zhang Y (2009) Protein structure prediction: when is it useful? Curr Opin Struct Biol 19: 145–155.
- 10. Petrey D, Honig B (2005) Protein structure prediction: inroads to biology. Mol Cell 20: 811–819.
- 11. Moult J, Fidelis K, Kryshtafovych A, Tramontano A (2011) Critical assessment of methods of protein structure prediction (CASP)–round IX. Proteins 79 Suppl 101–5.
- 12. Wu S, Zhang Y (2007) LOMETS: a local meta-threading-server for protein structure prediction. Nucleic Acids Res 35: 3375–3382.
- 13. Xu D, Zhang Y (2012) Ab initio protein structure assembly using continuous structure fragments and optimized knowledge-based force field. Proteins.
- 14. Berman HM, Westbrook J, Feng Z, Gilliland G, Bhat TN, et al. (2000) The Protein Data Bank. Nucleic Acids Res 28: 235–242.
- 15. Remmert M, Biegert A, Hauser A, Soding J (2012) HHblits: lightning-fast iterative protein sequence searching by HMM-HMM alignment. Nat Methods 9: 173–175.
- 16. 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. Nucleic Acids Res 25: 3389–3402.
- 17. Roche DB, Buenavista MT, Tetchner SJ, McGuffin LJ (2011) The IntFOLD server: an integrated web resource for protein fold recognition, 3D model quality assessment, intrinsic disorder prediction, domain prediction and ligand binding site prediction. Nucleic Acids Res 39: W171–176.
- 18. Yang Y, Faraggi E, Zhao H, Zhou Y (2011) Improving protein fold recognition and template-based modeling by employing probabilistic-based matching between predicted one-dimensional structural properties of query and corresponding native properties of templates. Bioinformatics 27: 2076–2082.
- 19. Peng J, Xu J (2010) Low-homology protein threading. Bioinformatics 26: i294–300.
- 20. Moult J, Fidelis K, Kryshtafovych A, Rost B, Tramontano A (2009) Critical assessment of methods of protein structure prediction - Round VIII. Proteins 77 Suppl 91–4.
- 21. Lundstrom J, Rychlewski L, Bujnicki J, Elofsson A (2001) Pcons: a neural-network-based consensus predictor that improves fold recognition. Protein Sci 10: 2354–2362.
- 22. Sali A, Blundell TL (1993) Comparative protein modelling by satisfaction of spatial restraints. J Mol Biol 234: 779–815.
- 23. Pandit SB, Zhang Y, Skolnick J (2006) TASSER-Lite: an automated tool for protein comparative modeling. Biophys J 91: 4180–4190.
- 24. Zhou H, Zhou Y (2005) SPARKS 2 and SP3 servers in CASP6. Proteins 61 Suppl 7152–156.
- 25. Wang G, Dunbrack RL Jr (2003) PISCES: a protein sequence culling server. Bioinformatics 19: 1589–1591.
- 26. 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.
- 27. Brenner SE, Koehl P, Levitt M (2000) The ASTRAL compendium for protein structure and sequence analysis. Nucleic Acids Res 28: 254–256.
- 28. Biegert A, Soding J (2009) Sequence context-specific profiles for homology searching. Proc Natl Acad Sci U S A 106: 3770–3775.
- 29. Sadreyev R, Grishin N (2003) COMPASS: a tool for comparison of multiple protein alignments with assessment of statistical significance. J Mol Biol 326: 317–336.
- 30. Soding J (2005) Protein homology detection by HMM-HMM comparison. Bioinformatics 21: 951–960.
- 31. Eddy SR (1998) Profile hidden Markov models. Bioinformatics 14: 755–763.
- 32. Bucher P, Karplus K, Moeri N, Hofmann K (1996) A flexible motif search technique based on generalized profiles. Comput Chem 20: 3–23.
- 33. Lobley A, Sadowski MI, Jones DT (2009) pGenTHREADER and pDomTHREADER: new methods for improved protein fold recognition and superfamily discrimination. Bioinformatics 25: 1761–1767.
- 34. Hughey R, Krogh A (1996) Hidden Markov models for sequence analysis: extension and analysis of the basic method. Comput Appl Biosci 12: 95–107.
- 35. Jones DT, Taylor WR, Thornton JM (1992) A new approach to protein fold recognition. Nature 358: 86–89.
- 36. Chang CC, Lin CJ (2011) LIBSVM: A library for support vector machines. ACM Transactions on Intelligent Systems and Technology 2: 27.
- 37. Zhang Y, Skolnick J (2004) Scoring function for automated assessment of protein structure template quality. Proteins 57: 702–710.
- 38. Pandit SB, Skolnick J (2008) Fr-TM-align: a new protein structural alignment method based on fragment alignments and the TM-score. BMC Bioinformatics 9: 531.
- 39. Needleman SB, Wunsch CD (1970) A general method applicable to the search for similarities in the amino acid sequence of two proteins. J Mol Biol 48: 443–453.
- 40. Skolnick J, Jaroszewski L, Kolinski A, Godzik A (1997) Derivation and testing of pair potentials for protein folding. When is the quasichemical approximation correct? Protein Sci 6: 676–688.
- 41. Rotkiewicz P, Skolnick J (2008) Fast procedure for reconstruction of full-atom protein models from reduced representations. J Comput Chem 29: 1460–1465.
- 42. MacKerell AD, Bashford D, Bellott M, Dunbrack RL, Evanseck JD, et al. (1998) All-Atom Empirical Potential for Molecular Modeling and Dynamics Studies of Proteins. J Phys Chem B 102: 3586–3616.
- 43. Xiang Z, Honig B (2001) Extending the accuracy limits of prediction for side-chain conformations. J Mol Biol 311: 421–430.
- 44. Joachims T (2006) Training Linear SVMs in Linear Time. Proceedings of the ACM Conference on Knowledge Discovery and Data Mining.
- 45. Zhang C, Liu S, Zhou H, Zhou Y (2004) An accurate, residue-level, pair potential of mean force for folding and binding based on the distance-scaled, ideal-gas reference state. Protein Sci 13: 400–411.
- 46. Jones DT (1999) Protein secondary structure prediction based on position-specific scoring matrices. J Mol Biol 292: 195–202.
- 47. Zhang Y, Skolnick J (2004) SPICKER: a clustering approach to identify near-native protein folds. J Comput Chem 25: 865–871.
- 48. Skolnick J, Zhou H, Brylinski M (2012) Further evidence for the likely completeness of the library of solved single domain protein structures. J Phys Chem B 116: 6654–6664.
- 49. Zhang Y, Skolnick J (2004) Tertiary structure predictions on a comprehensive benchmark of medium to large size proteins. Biophys J 87: 2647–2655.
- 50. Baldi P, Brunak S, Chauvin Y, Andersen CA, Nielsen H (2000) Assessing the accuracy of prediction algorithms for classification: an overview. Bioinformatics 16: 412–424.
- 51. Izarzugaza JM, Grana O, Tress ML, Valencia A, Clarke ND (2007) Assessment of intramolecular contact predictions for CASP7. Proteins 69 Suppl 8152–158.
- 52. Cheng J, Baldi P (2007) Improved residue contact prediction using support vector machines and a large feature set. BMC Bioinformatics 8: 113.
- 53. Randall A, Baldi P (2008) SELECTpro: effective protein model selection using a structure-based energy function resistant to BLUNDERs. BMC Struct Biol 8: 52.
- 54. Wang Z, Eickholt J, Cheng J (2011) APOLLO: a quality assessment service for single and multiple protein models. Bioinformatics 27: 1715–1716.
- 55. Zemla A (2003) LGA: A method for finding 3D similarities in protein structures. Nucleic Acids Res 31: 3370–3374.
- 56. Kabsch W (1976) A solution for the best rotation to relate two sets of vectors. Acta Cryst 32: 922–923.
- 57. Damm KL, Carlson HA (2006) Gaussian-weighted RMSD superposition of proteins: a structural comparison for flexible proteins and predicted protein structures. Biophys J 90: 4558–4573.
- 58. 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.
- 59. Zhang Y, Skolnick J (2004) Automated structure prediction of weakly homologous proteins on a genomic scale. Proc Natl Acad Sci U S A 101: 7594–7599.
- 60. Laskowski RA, MacArthur MW, Moss DS, Thornton JM (1993) PROCHECK: a program to check the stereochemical quality of protein structures. J Appl Cryst 26: 283–291.
- 61. Hartshorn MJ (2002) AstexViewer: a visualisation aid for structure-based drug design. J Comput Aided Mol Des 16: 871–881.
- 62. Ginalski K, Elofsson A, Fischer D, Rychlewski L (2003) 3D-Jury: a simple approach to improve protein structure predictions. Bioinformatics 19: 1015–1018.
- 63. Kurowski MA, Bujnicki JM (2003) GeneSilico protein structure prediction meta-server. Nucleic Acids Res 31: 3305–3307.
- 64. Wallner B, Larsson P, Elofsson A (2007) Pcons.net: protein structure prediction meta server. Nucleic Acids Res 35: W369–374.
- 65. Cheng J, Baldi P (2006) A machine learning information retrieval approach to protein fold recognition. Bioinformatics 22: 1456–1463.
- 66. Zhang Z, Kochhar S, Grigorov MG (2005) Descriptor-based protein remote homology identification. Protein Sci 14: 431–444.
- 67. Han S, Lee BC, Yu ST, Jeong CS, Lee S, et al. (2005) Fold recognition by combining profile-profile alignment and support vector machine. Bioinformatics 21: 2667–2673.
- 68. Rai BK, Fiser A (2006) Multiple mapping method: a novel approach to the sequence-to-structure alignment problem in comparative protein structure modeling. Proteins 63: 644–661.
- 69. Kosinski J, Gajda MJ, Cymerman IA, Kurowski MA, Pawlowski M, et al. (2005) FRankenstein becomes a cyborg: the automatic recombination and realignment of fold recognition models in CASP6. Proteins 61 Suppl 7106–113.
- 70. Gao X, Bu D, Xu J, Li M (2009) Improving consensus contact prediction via server correlation reduction. BMC Struct Biol 9: 28.
- 71. Punta M, Rost B (2005) Protein folding rates estimated from contact predictions. J Mol Biol 348: 507–512.
- 72. Xu J, Zhang Y (2010) How significant is a protein structure similarity with TM-score = 0.5? Bioinformatics 26: 889–895.
- 73. Rost B (1999) Twilight zone of protein sequence alignments. Protein Eng 12: 85–94.
- 74. Kmiecik S, Gront D, Kolinski A (2007) Towards the high-resolution protein structure prediction. Fast refinement of reduced models with all-atom force field. BMC Struct Biol 7: 43.
- 75. Dominy BN, Brooks CL (2002) Identifying native-like protein structures using physics-based potentials. J Comput Chem 23: 147–160.
- 76. Lee MR, Tsai J, Baker D, Kollman PA (2001) Molecular dynamics in the endgame of protein structure prediction. J Mol Biol 313: 417–430.
- 77. Jackson RM, Gabb HA, Sternberg MJ (1998) Rapid refinement of protein interfaces incorporating solvation: application to the docking problem. J Mol Biol 276: 265–285.
- 78. Trott O, Olson AJ (2010) AutoDock Vina: improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading. J Comput Chem 31: 455–461.
- 79. Tovchigrechko A, Wells CA, Vakser IA (2002) Docking of protein models. Protein Sci 11: 1888–1896.
- 80. Brylinski M, Skolnick J (2010) Q-Dock(LHM): Low-resolution refinement for ligand comparative modeling. J Comput Chem 31: 1093–1105.