• We are sorry, but NCBI web applications do not support your browser and may not function properly. More information
Logo of nihpaAbout Author manuscriptsSubmit a manuscriptNIH Public Access; Author Manuscript; Accepted for publication in peer reviewed journal;
Structure. Author manuscript; available in PMC Aug 10, 2012.
Published in final edited form as:
PMCID: PMC3154634
NIHMSID: NIHMS300419

Improving protein structure prediction using multiple sequence-based contact predictions

Summary

Although residue-residue contact maps dictate the topology of proteins, sequence-based ab initio contact predictions have been found little use in actual structure prediction due to the low accuracy. We developed a composite set of nine SVM-based contact predictors which are used in I-TASSER simulation in combination with sparse template contact restraints. When testing the strategy on 273 non-homologous targets, remarkable improvements of I-TASSER models were observed for both easy and hard targets, with P-value by student s t-test below 0.00001 and 0.001, respectively. In several cases, TM-score increases by >30%, which essentially converts “non-foldable” targets into “foldable” ones. In CASP9, I-TASSER employed ab initio contact predictions, and generated models for 26 FM targets with a GDT-score 16% and 44% higher than the second and third best servers from other groups, respectively. These findings demonstrate a new avenue to improve the accuracy of protein structure prediction especially for free-modeling targets.

Keywords: protein structure prediction, ab initio folding, contact prediction, threading

Introduction

The topology of protein three-dimensional structures can be specified by their inter-residue distance and contact maps. Thus, the structure of a protein molecule can be readily reconstructed by computer if all the native contacts are known. Using the power of contemporary protein structure prediction algorithms, which utilize various structural regularities such as predicted secondary structure and advanced force fields (Liwo et al., 1999; Roy et al., 2010; Sali and Blundell, 1993; Simons et al., 1997; Zhang and Skolnick, 2004a), the structure of a protein can be built based on just a small fraction of native contacts. For example, Li et. al. (Li et al., 2004) showed that only one native contact (from NMR nuclear Overhauser enhancement [NOE] data) for every eight residues is sufficient to guide the structure prediction tool TOUCHSTONE-II (Zhang et al., 2003) to construct a correct topology for single-domain proteins up to 200 residues. This is particularly encouraging because requiring fewer native contacts for structure reconstruction allows a significant reduction in experimental data collection efforts and thus makes possible the structure determination of a wide range of proteins for which obtaining a full set of native contacts is difficult.

For most proteins in nature, however, not even sparse experimental contact data are available, and the inter-residue contacts must be generated by computer-based predictions. Contact prediction methods can be largely classified into two types. The first is the template-based method (Misura et al., 2006; Shao and Bystroff, 2003; Skolnick et al., 2004; Wu and Zhang, 2007), i.e. collecting contacts from solved homologous proteins whose structures can be found in the PDB by sequence similarity search (Altschul et al., 1997) or threading algorithms (Bowie et al., 1991; Karplus et al., 1998; Soding, 2005; Wu and Zhang, 2008b). The accuracy of the contact prediction can be very high when closely homologous templates are identified, which has been shown to be extremely useful for high-resolution template-based protein structure prediction (Raman et al., 2009; Sali and Blundell, 1993; Zhang, 2009). Zhang et al. showed that contact predictions with an average accuracy >22% should have an overall positive effect on ab initio protein folding simulations (Zhang et al., 2003). However, one limitation of template-based contact prediction is that the accuracy highly depends on the availability of templates. For hard protein targets, i.e. those without homologous templates, template-based contact prediction usually has a low accuracy and therefore becomes useless for protein structure prediction.

The second type of contact prediction methods does not depend on protein template structures. Instead, contact predictions are derived from the primary sequence by identifying correlated mutations (Gobel et al., 1994) or machine learning methods (Cheng and Baldi, 2007; Shackelford and Karplus, 2007; Wu and Zhang, 2008a). For the hard free-modeling (FM) protein targets, it has been shown in the CASP experiments that the purely ab initio sequence-based contact predictions have a higher accuracy than those collected from the best template-based models (Ezkurdia et al., 2009). Despite its appealing feature of non-dependence on template structures, the usefulness of sequence-based contact prediction in 3D structure prediction has not yet been systematically assessed and demonstrated in the literature. Considering the still low accuracy (typically ~20-30%) of the state-of-the-art contact prediction algorithms (Cheng and Baldi, 2007; Shackelford and Karplus, 2007; Wu and Zhang, 2008a), it is particularly important to know when the ab initio contact predictions should be used, whether and how they should be combined with the template-based contact information, and how they can be best geared into the conventional template assembly algorithms such as I-TASSER (Roy et al., 2010; Wu et al., 2007), MODELLER (Sali and Blundell, 1993), and ROSETTA (Simons et al., 1997).

In this work, we aim to provide a systematic examination of these open questions in the framework of I-TASSER (Roy et al., 2010; Wu et al., 2007; Zhang and Skolnick, 2004a) which was designed to construct protein structures by assembling template structure fragments identified by threading (Wu and Zhang, 2007). To address the major weakness of high false positive rate in single sequence-based contact predictors (Cheng and Baldi, 2007; Shackelford and Karplus, 2007; Wu and Zhang, 2008a), we developed a composite set of 9 contact predictors, each trained on different atom types with different distance cutoffs, using support vector machines (SVMs). The combination of ab initio contacts with sparse template-based restraints in I-TASSER is carried out differently for easy and hard targets, and improvements are demonstrated for both groups. Notably, encouraging examples have been found where non-foldable targets can be converted into foldable ones owing to the use of ab initio contact predictions.

Results

The ab initio contact predictions are generated by an extended version of SVMSEQ (Wu and Zhang, 2008a), with individual predictors trained on contacts defined by Cα, Cβ atoms and side-chain centers of mass, each with 3 distance cutoffs (7, 8, 9 Å), as described in Methods. The contacts are then used as restraints in I-TASSER simulations (Wu et al., 2007). A total of 273 non-homologous proteins have been collected as our benchmark test proteins, which includes 253 proteins collected from the PDB by PISCES (Wang and Dunbrack, 2003) and 8 FM targets from CASP7 and 12 FM targets from CASP8. All the proteins are single domain proteins. Since I-TASSER starts with threading templates identified by LOMETS (Wu and Zhang, 2007), for fair testing, all templates having >30% sequence identity with the target were excluded from the LOMETS template library. As SVMSEQ was trained on 500 training proteins, to avoid over-training, we have confirmed that the benchmark proteins all have a sequence identity below 25% to the SVMSEQ training proteins.

The structures of the target proteins were predicted by I-TASSER, either with or without sequence-based predicted contacts. The accuracy of the models was evaluated using RMSD and TM-score (Zhang and Skolnick, 2004b). TM-score measures the similarity of two structures in a chain length independent way, which is more sensitive than RMSD to the topological similarity of structures especially when RMSD is high. TM-score has a value in the [0, 1] interval; a TM-score < 0.17 indicates a random similarity and a TM-score > 0.5 corresponds to protein structures having the same global fold as defined in SCOP and CATH (Xu and Zhang, 2010).

A systematic comparison of the Cα contact prediction by SVMSEQ with that by template-based predictions was conducted earlier (Wu and Zhang, 2008a). According to this study, the template-based contact prediction typically outperforms the sequence-based, ab initio contact prediction for “easy” and “medium” targets, i.e. when homologous templates are available. But for “hard” targets where no reliable templates can be identified, the accuracy of the SVMSEQ prediction is ~12-25% higher than that produced by LOMETS threading. It was also shown that the accuracy of the SVMSEQ prediction is close to or slightly higher than that of other state-of-the-art ab initio contact predictors, including SAM server (Shackelford and Karplus, 2007) and SVMCON (Cheng and Baldi, 2007). In this study, we will not repeat these comparisons and focus instead on finding the best combination of the ab initio and template-based contact predictions which can be optimally used in protein structure assembly for both hard and easy targets.

Sequence-based contact predictions used for hard targets

First, we test the usefulness of sequence-based contact predictions in protein structure prediction for hard targets. We selected 164 non-homologous proteins with lengths ranging from 41 to 207 residues which were classified as “hard” targets by LOMETS (Wu and Zhang, 2007) because the Z-scores of all threading templates are lower than the predefined thresholds, meaning that no threading program can identify a good template. The hard targets include 59 alpha, 22 beta and 83 α+/β proteins.

As shown in Table 1, the average accuracy (=17.9%) of the template-based contact predictions on the hard targets is low, compared to 22% that was found necessary to improve ab initio modeling (Zhang et al., 2003). In particular, for the long-range contacts (|i−j|>24), the Cα and side-chain contact accuracies are 13.1% and 18.0%, respectively. We therefore combined the contact predictions of LOMETS and SVMSEQ using a weighted sum of the confidence scores (see Methods). The combined contacts, which have a much higher accuracy (14.1% and 28.2%, respectively), were then used in I-TASSER as constraints in the structural assembly simulations.

Table 1
Contact prediction accuracy (ACC) and coverage (cov) for 164 hard targets.

The average TM-score of the first-ranking models is 0.386, which is consistent with the difficulty of structure prediction for these targets. For 36 out of the 164 hard targets, the first-ranking models have a good quality with TM-score >0.5, indicating successful prediction of the global fold. If the best of the top 5 models is considered for each target, the average TM-score increases to 0.410, and 41 targets have predicted models with TM-score >0.5.

When structures are predicted with the original I-TASSER procedure which only uses template-based distance and contact predictions (“old I-TASSER”), the average TM-score of the first-ranking models is 0.369. Thus, the “new” I-TASSER achieves a ~4.6% higher average TM-score than the “old” one. The P-value by the paired student s t-test for the two sets of models is below 0.001, which demonstrates that the TM-score improvement by SVMSEQ is statistically significant. Figure 1 shows a head-to-head comparison of the TM-scores obtained with and without ab initio contact predictions. There are a number of targets which shows significant TM-score increase. For example, there are 15 proteins which have a TM-score increase by >0.12, 10 proteins with a TM-score increase by >0.2, 6 proteins with a TM-score increase by >0.25. Most of these targets conduct a TM-score transition from far below 0.5 to above 0.5, indicating that the use of sequence-based contacts converted these targets from “nonfoldable” to “foldable” if we consider TM-score >0.5 as a quantitative criterion for assessing whether two protein structures have a similar fold (Xu and Zhang, 2010).

Figure 1
TM-score of the first-ranking models generated by the normal implementation of I-TASSER on 164 non-homologous Hard targets (stars) and 109 non-homologous Easy/Medium targets (circles) versus that of I-TASSER when using ab initio contact predictions.

On the contrary, there is only one protein, the Gamma subunit of the dissimilatory sulfite reductase (PDB ID: 1sauA), where the ab initio contact prediction reduces the TM-score of the I-TASSER model by >0.12. For this target, the SVMSEQ contact predictions in the N-terminal have low accuracy which distracted the N-tail (1P-17F) flip away from the core; this results in an overall TM-score deterioration from 0.513 to 0.386, a reduction of 0.127. However, the global topology in core region of the protein (39S-114V) keeps unchanged.

What is the reason for the improvement of the predicted structures? We take a detailed look at Table 1, which summarizes the accuracy (number of correct predictions divided by total number of predictions) and coverage (number of correct predictions divided by number of contacts in target) of contact predictions, comparing the consensus contacts obtained from combining the sequence- and template-based contacts with the template-based contacts from LOMETS only. When short-, medium- and long-range contacts are combined (|i−j|≥6), the average accuracy of the consensus contact predictions is 0.261 (with a coverage=14.1%) for the Cα−Cα contacts, which is 35% higher than that of the template-based contacts (accuracy=0.193, coverage=17.1%). For the contacts between side-chain centers, the average accuracy of the consensus contact predictions is 0.362 (coverage=14.4%), almost twice as much as that of the template-based ones (accuracy=0.194, coverage=20.7%). Combining the Cα and side-chain center contacts, the overall consensus contact predictions achieve an accuracy of 0.310 (coverage=13.2%), which is 73% more than that of the template-based ones (accuracy=0.179, coverage=23.1%).

Here, we want to mention that we did not perform a “fair” comparison of the two sets of contact predictions according to the conventional standard that compares the accuracies of predictions with the same coverage (Wu and Zhang, 2008a). However, we found that the structure prediction results were more sensitive to the correctness of contacts (the accuracy) than the number of predictions (the coverage) although both are important (Zhang, 2009). The balance of accuracy and coverage of the consensus contacts was optimized on an independent set of training proteins with the purpose of maximizing the TM-score of the final model (see Supplemental Information). In fact, even for a single set of contact predictions, the accuracy can be slightly increased by reducing the coverage because there must be a positive correlation between the confidence score and the accuracy of the prediction in any reasonable contact predictor. However, we have previously shown that ab initio contact prediction yields substantial novel contacts that are added to the template-based contacts, thus the accuracy increase attained by taking a consensus is significantly beyond that which can be achieved by simply reducing the coverage (Wu and Zhang, 2008a).

As shown in the following examples, most of the improvement of the final models, especially for the proteins with a >0.25 TM-score increase, is indeed due to the increase in contact prediction accuracy. As shown in Table 1, the accuracy of the short-range contacts is higher than that of the long-range ones. The improvement of the structural models, however, is mainly due to the long-range contacts. In fact, when we removed the long-range ab initio contacts, there was almost no difference between the average TM-scores of the final models generated by “old” and “new” I-TASSER. However, the overall improvement of the I-TASSER models when both long- and short-range contacts were used was more pronounced than with using only long-range contacts, indicating that the short-range contacts are still necessary for fine-tuning the packing of local structures.

Examples for successful prediction on hard targets

We now take a closer look at the targets where a striking improvement in model accuracy occurs. We choose three typical examples which all have a TM-score improvement >0.25. The first such protein is “1kafA”, the DNA binding domain of the phage T4 transcription factor MotA (Li et al., 2002). It is an α/β protein (3 α-helices and 6 β-strands) with 108 residues. If I-TASSER is used with only template-based contact predictions, the first-ranking model has a wrong fold with TM-score=0.329 and RMSD=12.6Å, as seen in Fig. 2B. The model is not improved much when compared with the best identified template (PDB ID 2rsl , chain A ) which has a TM-score=0.330 to “1kafA” as reported by the structural alignment program TM-align (Zhang and Skolnick, 2005). The structure of an N-terminal part, which is similar to the template, is correctly predicted but the remaining C-terminal segment is at a large distance from the native structure, where three helices are on the opposite side of the β-sheet. When the sequence-based contact predictions are used in I-TASSER, the C-terminal segments are drawn closer to native owing to the correctly predicted contacts between the helices and β-strands. This places the helices to the correct side of the β-sheets, and the TM-score of the first-ranking model increases to 0.590 with an RMSD=3.8Å (Figure 2C).

Figure 2
An illustrative example of I-TASSER modeling for the target protein “1kafA”. (A) Experimental structure. (B) Model generated by I-TASSER without using ab initio contact prediction. (C) Model generated by I-TASSER with ab initio contact ...

The accuracy of template-based contact predictions for this target is 0.31 (/0.21) and the coverage is 25% (/23%) for contacts between Cα atoms (/sidechain centers). When the ab initio contact predictions are used to take a consensus with the template-based contacts, the accuracy of the Cα contacts increases by 71%, to 0.53, and the accuracy of the side-chain contacts doubles (=0.48), although the overall coverage slightly reduces. Remarkably, the contacts between helices and β-strands were newly introduced by ab initio contact predictions which helped improve the overall topology.

Another reason for the improvement is the introduction of Cβ contact predictions. LOMETS does not include Cβ contacts, while the average accuracy of Cβ contacts predicted by SVMSEQ is ~0.78. These changes are reflected in the Cα and side-chain center contact maps as shown in Figs. 2D and 2E, respectively. The blue “+” symbols (consensus contact predictions) in the lower triangle are much more symmetrical to the green “*” symbols (representing the native contacts) in the upper triangle than the red “*” symbols (representing the template-based contact predictions), clearly showing that the consensus contacts have a higher accuracy than the purely template-based ones. The contact maps also show that by taking a consensus of the sequence- and template-based contact predictions, many wrong template-based contact predictions (false positives) are filtered out.

The second example is the target “1zkeA”, which is the HP1531 protein from Helicobacter pylori. Its function is unknown, and it consists of an 81-residue chain that folds into 3 α-helices, with the N- and C- terminal α-helices being nearly anti-parallel. If the old I-TASSER procedure with default contacts and intrinsic potential is used, the first-ranking model has a TM-score of 0.252 (RMSD=14.1Å), with an incorrect topology containing four bent helices as shown in Fig. 3B. After the introduction of ab initio contact predictions into I-TASSER, the first-ranking model has a correct topology (see Fig. 3C) with TM-score=0.591. The N- and C- terminal helices are now extended and correctly placed, although the middle helix still has some error which is mainly due to the incorrect secondary structural prediction (as a loop).

Figure 3
An illustrative example of I-TASSER modeling for the target protein “1zkeA”. (A) Experimental structure. (B) Model generated by I-TASSER without ab initio contact prediction. (C) Model generated by I-TASSER with ab initio contact prediction. ...

In this example, the improvement is not due to the increase in the accuracy of Cα contact predictions because the average accuracy of both the template-based and the consensus contact predictions is zero, although an one-residue shift in the ab initio contacts would result in correct long-range Cα contact, as seen in the Cα contact map in Fig. 3D. If we look at the side-chain center contact map in Fig. 3E, however, the template-based contacts scatter all over the triangle area and most are false-positive; these contacts drove the two long helices into four bent short helices during the I-TASSER assembly process. When a consensus of sequence-based and template-based contact predictions is taken, those false positives in off-diagonal lines of Fig. 3E are effectively filtered out. As a result, the average accuracy of the consensus side-chain center contact prediction doubles, from 0.20 to 0.46. Thus, the higher contact accuracy, i.e. the removal of noise, is the main reason why the model is greatly improved by the introduction of ab initio contact prediction. This example also highlights the necessity of multiple contact predictors because the single Cα SVMSEQ predictor does not help improve the modeling accuracy.

The third example is from the target “1zv1A”, the dimerization domain of the doublesex protein from Drosophila melanogaster (Bayrer et al., 2005). It has 59 residues and its structure consists of three α-helices. The “old” I-TASSER generates a first-ranking model with a TM-score=0.454 and an RMSD=5.4Å, which is a considerable improvement compared to the best identified template, the second domain of ”1vdu”, chain A, which has a TM-score=0.302 to “1zv1A” as reported by TM-align (Zhang and Skolnick, 2005). Nevertheless, the orientation of the N-terminal and middle α-helices in the I-TASSER model is incorrect. The three helices are packed in an approximately parallel and anti-parallel bundle without exhibiting the subtle tilt in the native structure (Fig. 4B). When the ab initio contacts are introduced into I-TASSER, the quality of the final model is improved further, with the TM-score increasing to 0.592 and RMSD reduced to 3.8 Å (Fig. 4C). In this example, the intrinsic potential of I-TASSER tends to pack the helices into a more compact core with the tilted helix orientation but the strong contacts collected from template structures pushes the helices in the orientation seen in the templates. As shown in Figs. 4D and 4E, the incorporation of the ab initio contact predictions eliminated most of the false-positive contacts from the templates and doubled the accuracy of the whole set of contact restraints, which resulted in an adjustment of the relative orientation of the three helices.

Figure 4
An illustrative example of I-TASSER modeling for the target protein “1zc1A”. (A) Experimental structure. (B) Model generated by I-TASSER without ab initio contact prediction. (C) Model generated by I-TASSER with ab initio contact prediction. ...

Sequence-based contact predictions used for easy/medium targets

Although our focus was to use ab initio contact prediction to improve the structure prediction of hard targets, we have occasionally observed in our benchmark tests that the contact predictions can also have a positive effect for easy and medium targets. Here we conduct a systematic examination of the possible usefulness of ab initio contact prediction for easy and medium targets, i.e. where templates with a high Z-score are identified in threading (Wu and Zhang, 2007). Because the contacts derived from such templates usually have a higher accuracy than those generated by sequence-based methods (Wu and Zhang, 2008a), we expect that the latter becomes most useful for targets with substantial weakly aligned or unaligned regions. We will focus our analysis on these cases. Because of the imbalance of the contact accuracies of the ab initio and template-based predictions, instead of taking the consensus of different contact predictions, here we implement all contact predictions as separate restraints in I-TASSER (see Methods). One advantage of this approach is that the restraints for the ab initio contacts will be in effect in the threading-unaligned regions, where they would otherwise be filtered out if the consensus method were used.

We collected a set of 109 targets which had been categorized by LOMETS as easy/medium targets but have >10% regions not aligned by threading; the sequence-based predictions on these proteins generated >15% new contacts that do not appear among the template-based contact predictions (Wu and Zhang, 2007). This set of proteins includes 24 alpha, 11 beta, and 74 α+/β proteins, with lengths ranging from 31 to 273 residues. As with the hard targets, we used I-TASSER to generate models with and without including the sequence-based contacts.

For the 109 easy/medium targets, the average TM-score of the first-ranking I-TASSER models is 0.714 with the inclusion of the sequence-based contacts, 2.7% higher than that without including them (0.695). Because for easy/medium targets, the template-based contact predictions are usually more accurate than the sequence-based ones (Wu and Zhang, 2008a), it is not surprising that adding the sequence-based contacts yields slightly less improvement in the overall topology than that for hard targets. But clearly, it does no harm even if the sequence-based contact prediction has an obviously higher false-positive rate. Actually, the paired student s t-test P-value of the two sets of models is 9.3×10−6, which is statistically even more significant than what we observed in the hard proteins (P-value=0.00091). This is mainly because of the fact that in the well-aligned regions where the template-based restraints from consensus threading alignments are strong and dominant, the I-TASSER simulation is not influenced by the SVMSEQ predictions which are relatively more divergent. But in the regions where threading has low confidence, the template-based restraints are usually divergent and weak, and ab initio contact predictions become dominant, which helps in improving the modeling accuracy due to the higher accuracy of predictions in these regions. Thus, overall, there are more proteins in the easy target achieving a positive TM-score improvement, which resulted in a lower P-value, although the average magnitude of improvement is not as big as that in the hard targets.

To illustrate the improvement in easy/medium proteins, we take the example of “T0437” from the blind CASP8 experiment, where we tested I-TASSER with the ab initio contact predictions for the first time. This target is the yiiS protein from Shigella flexneri, which was categorized by CASP8 as a TMB-HA (Template Based Modeling - High Accuracy) target. In addition to an unstructured N-terminus (which was ignored in the CASP8 analysis), it contains 68 residues with 2 α-helices and 3 β-strands. The LOMETS threading results were dominated by the template “2jz5A”, for which the best threading alignment generated by HHsearch (Soding, 2005) has an RMSD =2.30 Å and TM-score =0.778. If we structurally align “2jz5A” to the experimental structure by TM-align, the RMSD is 1.34 Å with TM-score=0.838 (Figure 5A). Although the global topology of “2jz5A” matches the target well, there is a major mismatch in the region V49-T60 (the lower part of the second β-sheet). Correspondingly, there is no correct contact prediction from LOMETS in this region (Figure 5B). The sequence-based SVMSEQ contact prediction, however, generates 10 correct Cα contact predictions in this region (2 others are false positive, Figure 5B). These restraints helped I-TASSER generate models with a correct β-sheet structure in this region. The RMSD of the overall model is 1.13 Å, which is even closer to native than the best structural alignment (Figure 5C). In this example, although the overall accuracy of the SVMSEQ prediction is still lower than that from LOMETS, the novel contacts from the ab initio prediction improve the quality of local structures. In other regions (e.g. the N-terminal β-sheet), SVMSEQ generates a number of false positive contact predictions. Since the LOMETS predictions provide strong consensus restraints, these weak false-positive predictions did not reduce the modeling accuracy in those regions.

Figure 5
Analysis of I-TASSER modeling for the CASP8 “easy” target “T0437”. (A) The best template protein “2jz5A” (top panel) and its optimal structural alignment by TM-align (Zhang and Skolnick, 2005) on the experimental ...

Modeling of hard targets in CASP9 experiment

The SVMSEQ contact predictions were also used to assist I-TASSER modeling in the CASP9 experiments. According to the assessor s classification, there were 26 proteins/domains which had no similar structures in the PDB and belonged to the Free Modeling (FM) targets. In Table S1, we present a summary of the automated I-TASSER predictions, together with 19 best servers from other groups, on the 26 FM targets/domains (T0529_1, T0531_1, T0534_1, T0534_2, T0537_1, T0537_2, T0544_1, T0547_3, T0547_4, T0550_1, T0550_2, T0553_1, T0553_2, T0561_1, T0571_1, T0571_2, T0578_1, T0581_1, T0604_1, T0604_3, T0608_1, T0616_1, T0618_1, T0621_1, T0624_1, and T0629_2), which have lengths ranging from 56 to 333 residues.

The accuracy of SVMSEQ contact prediction is highly correlated with the confidence score. For example, a Cα contact prediction with a SVMSEQ confidence score >0.8 is almost always correct and that with a confidence score >0.7 has an average accuracy 80%. However, for most hard targets, there may be very few predictions with high confidence score. For the FM target in CASP9, to cover sufficient contact predictions, we use the top 0.6*L (L is the length of proteins) contacts regardless of the accuracy or more if the confidence score of additional contacts is higher than the specific confidence score cutoffs (see SI). The average accuracy of the contact predictions by SVMSEQ is 27.6% with an average number of predictions =0.606*L for the 26 FM targets. The side-chain contacts collected from the LOMETS templates have an accuracy of only 11.9%, which confirms that the threading templates are poor for this protein set. According to the assessor s assessment (Grishin, 2010) (see also http://prodata.swmed.edu/CASP9/evaluation/domainscore_sum/human_server-best-Z.html), the total GDT-TS score of the I-TASSER server models is 39.86, which is 16% higher than the second best server and 44% higher than the third best server (Table S1).

In Figure 6, we present two representative examples: one is an alpha- and one is a beta-protein. “T0553” is the PBS linker domain from Anabaena sp. and is 141 residues long. Although the assessor split the target into two domains (“T0553_1” and “T0553_2”, both being FM domains), we analyze the target here as the whole chain because the I-TASSER server did model it as single domain and the experimental structure looks overall well packed (see top-middle panel in Figure 6). The SVMSEQ Cα contact prediction has only 6 contacts that have a confidence score above the threshold (see Supplemental Information) but all are correct. Regardless of the confidence threshold, the I-TASSER server used the top 85 (0.6*L) contact predictions from SVMSEQ, out of which 32 were correct, distributed quite evenly along the chain except for the second helix (I27-E45, see top-right panel of Figure 6). The other 8 SVMSEQ predictors have a comparable accuracy and coverage. Finally, the I-TASSER server built a model with TM-score=0.493 that is about 20% higher than the best prediction from all other servers. The TM-score of the best threading template “1k94A” identified by LOMETS is only 0.289.

Figure 6
I-TASSER modeling on two CASP9 hard targets: “T0553” (top panel) and “T0604_1” (bottom panel). The left panels are backbones of the native structures with cross line representing SVMSEQ Cα contact predictions at ...

Target “T0604_1” (M11-P86) is the N-terminal domain of the VP0956 protein from Vibrio parahaemolyticus. The I-TASSER server identified the M1-A102 stretch as a domain based on the LOMETS alignments. SVMSEQ generated 48 contact predictions out of which 35 were correct, resulting in an accuracy=72.9%, which is significantly higher than the accuracy of contacts from LOMETS (15.2%). In particular, most of the SVMSEQ contact predictions along the two β-sheets are correct (bottom-left panel of Figure 6), which drove the I-TASSER simulation to precisely pack the 3 β-strands. As a result, the first-ranking I-TASSER model has a TM-score=0.691 and RMSD=2.66 Å, while the best template for this domain identified by LOMETS is “3goaA” which only has a TM-score=0.345. In both cases of T0553 and T0604_1, the high accuracy of the composite contact predictions by SVMSEQ is essential to the success in the modeling.

Discussion

Residue-residue contacts predicted purely from protein sequence have been widely regarded as being of little use in protein structure prediction due to their low accuracy, especially compared to contacts from template structures. However, the low accuracy does not imply that sequence-based contacts are useless when appropriately combined with template-derived contacts. In this study, we present ways to combine residue contacts predicted from sequence with those predicted from threading templates, and demonstrate that employing sequence-based contact predictions does improve the accuracy of the models obtained from protein structure prediction.

Using I-TASSER as an illustrative framework, we designed new contact energy terms that allow introduction of sequence-based contact predictions in the energy function of Monte Carlo simulations. The way we modified the energy function to allow for the sequence-based contacts is, however, different for hard targets and medium/easy targets. For hard targets, to reduce the negative effect of false positive contact predictions, we take a consensus of the sequence- and template-based predicted contacts so that contacts that do not have sufficient combined confidence are eliminated. The consensus method takes a weighted average of the confidence scores of predicted contacts from nine different sequence-based contacts (generated by an extended version of SVMSEQ) and two sets of template-based contacts (those for Cα and those for sidechain centers), and then uses the consensus contacts having a confidence score larger than a threshold in the I-TASSER s energy function. This solution introduces a “filtering” effect that can remove bad template-based contacts. Applying this method to the test proteins, we find that for a number of cases, I-TASSER could successfully convert a non-foldable target with TM-score far below 0.5 to a foldable one with TM-score >0.5. The overall TM-score improvement by SVMSEQ is statistically significant with the P-value in student s t-test below 1.0×10−3. An analysis of the CASP9 blind test performed on 26 “Free Modeling” targets also demonstrates the significant value of the method in the structural modeling of hard targets.

The basis for taking a consensus of the contacts predicted in different ways is that the accuracy of the template-derived and the ab initio predicted contacts is comparable for the hard targets. For easy and medium targets, however, the accuracy of the template-based contacts is, for most protein regions, higher than that of sequence-based ab initio predictions and taking a consensus might therefore significantly degrade the overall accuracy of template-based predictions. To take advantage of the ab initio contact predictions, which are mainly useful in the weakly aligned or unaligned regions for the easy/medium targets, we incorporated both sets of contact predictions into the I-TASSER assembly simulation. This way, the highly accurate template-based predictions are assigned strong weights in the well-aligned regions due to their high confidence, and can automatically eliminate the negative influence of the sparse ab initio contact predictions. But in regions where template alignments are not available, sequence-based contacts become dominant and come to the rescue. Thus, the introduction of sequence-based contacts does not harm the modeling of structural regions which are sufficiently covered by template-based alignments, but is beneficial for the regions not covered by template-derived contacts. The overall TM-score improvement by SVMSEQ is shown to be statistically significant with the P-value in student s t-test below 1.0×10−5 for the easy/medium proteins. A successful example is the easy target “T0437” in CASP8, where, using sequence-based ab initio contact predictions, the model generated by the I-TASSER server had a high accuracy (RMSD=1.1 Å), which is closer to the native structure than even the best template in the best structural alignment.

In summary, although the accuracy of the ab initio contact prediction is generally low, it can still be used in protein structure assembly because it often complements the template-derived contacts in a way that eventually improves model accuracy. For hard targets, even though some weak templates may often be available, their number is small, and they are too distant from the target in most cases and thus may provide incorrect contact predictions. The use of sequence-based contacts, which are generated after learning from a large set of protein structures rather than a small number of possibly wrong templates, helps eliminate the false structural information coming from the templates. In the case of easier targets, there may be some regions that are not sufficiently covered by template-based contacts. Sequence-based contacts are helpful in the prediction of those regions.

Compared with others and our previous (less successful) attempts, the success of the procedure here is partly due to the fact that we generated 9 different sets of sequence-based contact predictions, which are designed to capture the contact maps defined using various distance cutoffs and various objective atoms; the larger number and the diversity of the generated contacts allows the more reliable contacts to be selected or weighted. Meanwhile, the variation of contact predictions and integration strategies gives us a variety of options and parameter sets to optimize our approaches while keeping the training and testing proteins strictly non-redundant. However, because we generated all sets by the same SVM algorithm and based on the same training set (500 non-redundant proteins), the diversity of the contacts is probably not as large as it could be. Using a broader variety of contacts (for example, from methods relying on evolutionary information from correlated mutations (Gobel et al., 1994; Halperin et al., 2006; Kundrotas and Alexov, 2006; Olmea and Valencia, 1997; Vicatos et al., 2005) or generated by other machine learning methods such as neural networks (Shackelford and Karplus, 2007; Tegge et al., 2009) would probably further improve the performance of the method. Finally, although the data in this work have been presented using I-TASSER as a framework, we expect that the method can be demonstrated as a general approach to improve the accuracy of protein structure prediction in many other template-based modeling algorithms including MODELLER (Sali and Blundell, 1993), ROSETTA (Simons et al., 1997) and TASSER (Zhang and Skolnick, 2004a).

Methods

A detailed description of the methods is provided in the Supplemental Information (SI). Here, we provide a short summary.

The contact energy used in original I-TASSER is defined as

Econtact_temp=w1(i,j}list1,ji+6f(dcaij)+w2(i,j)list2,ji+6g(dsgij)
(1)

where f(.) is a contact energy term encouraging satisfaction of Cα contact restraints (distance cutoff dcaij=6), g(·) is a contact energy term penalizing violation of side-chain contact restraints (distance cutoff dsgij=cut(A,B), see SI); list1 and list2 are predicted template-based Cα and side-chain center contact lists, respectively; w1 and w2 are weighting factors.

The sequence-based contact predictions are generated by extended versions of SVMSEQ (Wu and Zhang, 2008a) which were trained on Cα, Cβ atoms and side-chain centers of mass with three different distance cutoffs (a total of 9 types of SVMSEQ predictions). For hard targets, we first combine the sequence-based (SVMSEQ) and template-based (from LOMETS) contact predictions by taking a weighted average of their confidence scores:

conf(i,j)=n=110wnconfn(i,j),
(2)

where conf(i, j) is the consensus contact confidence score for residues i and j, confn(i, j) is the contact confidence score for the nth individual predictor (9 predictors are sequence-based, and the last one is template-based), and wn is the weighting factor for the nth predictor. With the new sets of consensus contacts, we then use Eq. 1 to apply contact restraints in the new I-TASSER simulation.

For easy and medium targets, since the template-based contact predictions have usually a higher accuracy than SVMSEQ, we do not construct a new set of contacts to replace the template-based ones. Instead, we add terms to the contact energy function of Eq. 1 corresponding to the nine sets of sequence-based contacts, i.e.

Econtact_consensus=Econtact_temp+k=19wk(i,j)listk,ji+6f(dkij),
[3]

Here, the same weight is used for all sequence-based contact predictors. In this way, the contacts predicted by a larger number of different predictors will naturally obtain a higher weight than those predicted by fewer predictors.

Supplementary Material

Acknowledgments

The project is supported in part by the Alfred P. Sloan Foundation, NSF Career Award (DBI 1027394), and the National Institute of General Medical Sciences (GM083107, GM084222). A. S. was supported by a grant from the Hungarian Scientific Research Fund (OTKA PD73096).

Footnotes

Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final citable form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

References

  • Altschul SF, Madden TL, Schaffer 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. [PMC free article] [PubMed]
  • Bayrer JR, Zhang W, Weiss MA. Dimerization of doublesex is mediated by a cryptic ubiquitin-associated domain fold: implications for sex-specific gene regulation. The Journal of biological chemistry. 2005;280:32989–32996. [PubMed]
  • Bowie JU, Luthy R, Eisenberg D. A method to identify protein sequences that fold into a known three-dimensional structure. Science. 1991;253:164–170. [PubMed]
  • Cheng J, Baldi P. Improved residue contact prediction using support vector machines and a large feature set. BMC Bioinformatics. 2007;8:113. [PMC free article] [PubMed]
  • Ezkurdia I, Grana O, Izarzugaza JM, Tress ML. Assessment of domain boundary predictions and the prediction of intramolecular contacts in CASP8. Proteins. 2009;77(Suppl 9):196–209. [PubMed]
  • Gobel U, Sander C, Schneider R, Valencia A. Correlated mutations and residue contacts in proteins. Proteins. 1994;18:309–317. [PubMed]
  • Grishin NV. Assessment of Free Modeling in CASP9 Experiment. The 9th Community Wide Experiment on the Critical Assessment of Techniques for Protein Structure Prediction (Asilomar Conference Grounds); Pacific Grove, CA. 2010.
  • Halperin I, Wolfson H, Nussinov R. Correlated mutations: advances and limitations. A study on fusion proteins and on the Cohesin-Dockerin families. Proteins. 2006;63:832–845. [PubMed]
  • Karplus K, Barrett C, Hughey R. Hidden Markov models for detecting remote protein homologies. Bioinformatics. 1998;14:846–856. [PubMed]
  • Kundrotas PJ, Alexov EG. Predicting residue contacts using pragmatic correlated mutations method: reducing the false positives. BMC Bioinformatics. 2006;7:503. [PMC free article] [PubMed]
  • Li N, Sickmier EA, Zhang R, Joachimiak A, White SW. The MotA transcription factor from bacteriophage T4 contains a novel DNA-binding domain: the 'double wing' motif. Molecular microbiology. 2002;43:1079–1088. [PubMed]
  • Li W, Zhang Y, Skolnick J. Application of sparse NMR restraints to large-scale protein structure prediction. Biophys J. 2004;87:1241–1248. [PMC free article] [PubMed]
  • Liwo A, Lee J, Ripoll DR, Pillardy J, Scheraga HA. Protein structure prediction by global optimization of a potential energy function. Proc Natl Acad Sci U S A. 1999;96:5482–5485. [PMC free article] [PubMed]
  • Misura KM, Chivian D, Rohl CA, Kim DE, Baker D. Physically realistic homology models built with ROSETTA can be more accurate than their templates. Proc Natl Acad Sci U S A. 2006;103:5361–5366. [PMC free article] [PubMed]
  • Olmea O, Valencia A. Improving contact predictions by the combination of correlated mutations and other sources of sequence information. Fold Des. 1997;2:S25–32. [PubMed]
  • Raman S, Vernon R, Thompson J, Tyka M, Sadreyev R, Pei J, Kim D, Kellogg E, DiMaio F, Lange O, et al. Structure prediction for CASP8 with all-atom refinement using Rosetta. Proteins. 2009;77(Suppl 9):89–99. [PMC free article] [PubMed]
  • Roy A, Kucukural A, Zhang Y. I-TASSER: a unified platform for automated protein structure and function prediction. Nat Protoc. 2010;5:725–738. [PMC free article] [PubMed]
  • Sali A, Blundell TL. Comparative protein modelling by satisfaction of spatial restraints. J Mol Biol. 1993;234:779–815. [PubMed]
  • Shackelford G, Karplus K. Contact prediction using mutual information and neural nets. Proteins. 2007;69:159–164. [PubMed]
  • Shao Y, Bystroff C. Predicting interresidue contacts using templates and pathways. Proteins. 2003;53(Suppl 6):497–502. [PubMed]
  • Simons KT, Kooperberg C, Huang E, Baker D. Assembly of protein tertiary structures from fragments with similar local sequences using simulated annealing and Bayesian scoring functions. J Mol Biol. 1997;268:209–225. [PubMed]
  • Skolnick J, Kihara D, Zhang Y. Development and large scale benchmark testing of the PROSPECTOR_3 threading algorithm. Proteins. 2004;56:502–518. [PubMed]
  • Soding J. Protein homology detection by HMM-HMM comparison. Bioinformatics. 2005;21:951–960. [PubMed]
  • Tegge AN, Wang Z, Eickholt J, Cheng J. NNcon: improved protein contact map prediction using 2D-recursive neural networks. Nucleic Acids Res. 2009;37:W515–518. [PMC free article] [PubMed]
  • Vicatos S, Reddy BVB, Kaznessis Y. Prediction of distant residue contacts with the use of evolutionary information. Proteins-Structure Function and Bioinformatics. 2005;58:935–949. [PubMed]
  • Wang G, Dunbrack RL., Jr PISCES: a protein sequence culling server. Bioinformatics. 2003;19:1589–1591. [PubMed]
  • Wu S, Skolnick J, Zhang Y. Ab initio modeling of small proteins by iterative TASSER simulations. BMC Biol. 2007;5:17. [PMC free article] [PubMed]
  • Wu S, Zhang Y. LOMETS: a local meta-threading-server for protein structure prediction. Nucleic Acids Res. 2007;35:3375–3382. [PMC free article] [PubMed]
  • Wu S, Zhang Y. A comprehensive assessment of sequence-based and template-based methods for protein contact prediction. Bioinformatics (Oxford, England) 2008a;24:924–931. [PMC free article] [PubMed]
  • Wu S, Zhang Y. MUSTER: Improving protein sequence profile-profile alignments by using multiple sources of structure information. Proteins. 2008b;72:547–556. [PMC free article] [PubMed]
  • Xu J, Zhang Y. How significant is a protein structure similarity with TM-score = 0.5? Bioinformatics. 2010;26:889–895. [PMC free article] [PubMed]
  • Zhang Y. I-TASSER: Fully automated protein structure prediction in CASP8. Proteins. 2009;77:100–113. [PMC free article] [PubMed]
  • Zhang Y, Kolinski A, Skolnick J. TOUCHSTONE II: A new approach to ab initio protein structure prediction. Biophys J. 2003;85:1145–1164. [PMC free article] [PubMed]
  • Zhang Y, Skolnick J. Automated structure prediction of weakly homologous proteins on a genomic scale. Proc Natl Acad Sci USA. 2004a;101:7594–7599. [PMC free article] [PubMed]
  • Zhang Y, Skolnick J. Scoring function for automated assessment of protein structure template quality. Proteins. 2004b;57:702–710. [PubMed]
  • Zhang Y, Skolnick J. TM-align: a protein structure alignment algorithm based on the TM-score. Nucleic Acids Research. 2005;33:2302–2309. [PMC free article] [PubMed]
PubReader format: click here to try

Formats:

Related citations in PubMed

See reviews...See all...

Cited by other articles in PMC

Links

Recent Activity

Your browsing activity is empty.

Activity recording is turned off.

Turn recording back on

See more...