Blind prediction of quaternary structures of homo-oligomeric proteins from amino acid sequences based on templates

Background : Prediction of protein tertiary and quaternary structures helps us to understand protein functionality. While tertiary structure prediction techniques have been much improved over the last two decades, quaternary structure (homo-oligomer) prediction has not been paid much attention to. Results : We show the results of the assessment of our simple auto server prediction and manual prediction of protein quaternary structure from its amino acid sequence based on templates. They were tested in the 9th Critical Assessment of Protein Structure Prediction (CASP9) experiment. CASP experiments are the only true blind test for protein tertiary and quatenary structure prediction from amino acid sequence alone and therefore they are the most severe tests in the field of protein stucture prediction. Our simple auto server prediction could generate successful models for 14 out of 58 targets. Human experts could generate successful models for 11 out of 16 targets and most of them were better than those by our auto server. Conclusions : The results show the efficiency of our template-based protein quaternary structure prediction approaches and provide useful information for improvement of the accuracy of template-based quaternary structure prediction.


& Background
Proteins are capable of self-assembly and form homo-oligomers or homo-multimers. It has been estimated that about half of the known proteins are homo-oligomers [1,2]. Although the practical biological reasons that induce self-association of proteins are not yet completely understood, various examples of relationships between formation of homo-oligomers and their functions have been identified [3]. Catalytic sites of enzymes and ligand-binding sites of receptors are frequently located at the interface of subunits in order to form large binding pockets (Morita et al., unpublished results). Multiple binding sites formed in a given homo-oligomer can increase binding affinity between the protein and its ligand in a phenomenon known as the "multivalent effect" [4][5][6]. Certain membrane proteins control signal transduction by transient dimerization [7,8]. Consequently, defects in oligomeric states can cause diseases, and control of oligomeric states may provide a means for treating human diseases [9].
Thus, prediction of protein quaternary structures as well as tertiary structures is expected to shed light on biological issues. Over the last two decades, protein tertiary structure prediction techniques from their amino acid sequences have been significantly improved based on the development of novel prediction techniques and enlargement of protein structure databases. A number of websites and tools for protein tertiary structure prediction have been developed. On the other hand, few methods for predicting protein quaternary structures from their amino acid sequences have been reported so far. Baker and co-workers have extended their tertiary structure prediction framework ROSETTA to predict quaternary structures of symmetrical oligomeric proteins. This method is applicable when the symmetry type of a protein is known a priori [10]. Chen and Skolnick developed M-TASSER to predict dimeric structures from amino acid sequences [11]. Protein-protein docking tools can also be applied for the prediction of a quaternary structure of a homo-oligomeric protein with or without extensions [12]. These methods are aimed at de novo prediction (prediction without templates).
On another front, there are currently many templates for protein structure prediction available from the Protein Data Bank (PDB) [13], and such templates are constantly increasing in number. For tertiary structure prediction, model structures by using template-based modeling are so accurate that they can be applied for molecular replacement in many cases. Recently, Krissinel and Henrick developed PISA (Protein Interactions, Surface, and Assembly) to detect the most probable homo-oligomeric states of proteins solved by X-ray diffraction [14]. The success rate of PISA is 80-90% according to the authors' benchmark results. These advances have made the concept of template-based quaternary structure modeling a practical reality.
In this paper, we describe our template-based protein quaternary (homo-oligomer) structure prediction methods from amino acid sequences, which were tested in the 9 th Critical Assessment of Protein Structure Prediction (CASP9) experiment [15]. CASP experiments are the only true blind test for protein tertiary and quaternary structure prediction from amino acid sequences alone, in that various groups predict structures of proteins whose coordinates are not yet publically available. Usually, the performance of a prediction technique is estimated by the prediction results for a set of proteins whose coordinates are publically available. Though information about their structures are of course concealed during the test, there is room for overfitting of prediction parameters, which CASP experiments do not doi: 10.7243/2050-2273-1-1 have in principle. Actually, we often experience that performance of our developed prediction tools for CASP experiments are worse than we expect. So CASP experiments are the most severe tests in the field of protein structure prediction. In CASP9, quaternary structure prediction was assessed for more than 20 targets for prediction by auto-servers and more than 50 targets for manual prediction. Models from 7 auto servers and 16 human groups were submitted. Although there were a few targets for hetero-dimer predictions using sequence information in Critical Assessment of Prediction of Interactions (CAPRI) experiment [16], this is the first systematic assessment of the blind prediction of protein homooligomer structures for a number of targets. This requires to predict not only quaternary structures but also oligomeric states.
Our methods based on templates were ranked first among all quaternary structure predictors in both the auto server category and the human prediction category in CASP9 [15]. Our results demonstrate the effectiveness of the concept of template-based protein quaternary structure modeling.

Methods
We have tried 2 approaches for predicting protein quaternary structures: prediction by an auto server and prediction by a manual method utilizing the results of our another server predicting tertiary structures. Protein structures to be predicted (called "targets") in CASP9 were grouped into 2 categories: targets designated only for server prediction (denoted as "server-only" targets) and targets for both human and server prediction (denoted as "human/server" targets). Auto servers were supposed to make predictions for all targets, while human predictors only made predictions for selected human/server targets. Auto servers and human predictors were required to submit their predictions from amino acid sequences within 72 hours and 3 weeks, respectively. For human predictors, the number of chains in the native oligomer was occasionally available (Table S1).

Auto server:
The tertiary structure (monomer) prediction of a target protein was performed using a combination of template-based modeling and a consensus-based model quality assessment. Template structures were detected by widely used template detection tools: PDB-BLAST [17], FUGUE [18], and HHsearch [19]. Sequence-structure alignments were generated by T-COFFEE [20] and our pairwise alignment tool REALIZE, which employs an environment dependent position-specific gap penalty. About 300∼3000 models per target were generated from these alignments using MODELLER [21]. One model was selected according to a consensus-based model quality assessment approach. For this quality assessment, distances between all-by-all model structures were calculated using an S-score [22] per residue and the model with minimum averaged distances was selected based on the heuristics defining the most probable model as being the closest to the center of the generated models [23]. The quaternary structure of the selected monomer model was deduced from those of the template proteins used in tertiary structure modeling. The quaternary structures of the template proteins were obtained from the PISA server [http://pdbe.org/pisa/] [14]. The quaternary structure of the selected model was decided upon by the majority of the quaternary structures of those templates (i.e. voting). Here, only templates with TM-scores [24], a measure of structural similarity (from 0 to 1 and 1 means perfect match), greater than 0.3 with the selected monomer model were used. Finally, the quaternary structure of the target protein was obtained by superimposing the monomer structure of the selected model on the quaternary structure of the template protein with the highest TM-score with respect to the selected model. Structural refinements to avoid collisions of atoms around interfaces and application of distance constraints between subunits were not performed because of the prediction time limitation and our computational resources. Although refinement to avoid collisions is important for precise quaternary structure prediction, we wanted to observe the ability of this rather simple automated prediction to detect interfaces between chains in oligomeric states.

Manual method
For human predictions, human experts made modifications to our server prediction. Selections of template proteins and quaternary structures were reexamined by human experts with particular attention to select the most probable number of chains in the quaternary structure. When a proper template could not be obtained for building reasonable template-based models, quaternary structure modeling was not tried. For the other cases, templates which were suitable for quaternary structure prediction rather than for tertiary structure prediction were selected here, e.g., templates which have good alignments in the interface regions between different chains were used. The sequence alignments between the target protein and the selected templates were also manually corrected according mainly to position-specific sequence profile, secondary structure, and environment dependent position-specific gap penalty. Sequence alignments in not only hydrophobic core region but also inter-chain interface region were taken more care of. Model building with MODELLER using oligomeric state templates was also performed manually considering to avoid collisions around the interface regions.

Performance evaluation
The performance of our predictions was assessed according to the quaternary structures given in REMARK 350 records of the PDB files. An author-determined biological unit was used when it was provided. Otherwise, a software-determined quaternary structure, which appeared first in the PDB header, was used. We evaluated prediction accuracy according to the consistency of the number of subunits, "contact agreement score (S agree )," which was used in the official assessment of CASP9, and overall TM-score calculated by MM-align [25], which we have called "MM-score" in this paper to avoid confusion. The S agree score represents the structural similarity at interfaces between chains. S agree is defined as fraction of correctly predicted contacts between subunits and is calculated by the following equations: where residue i of 1 subunit and residue j of another subunit are regarded as being "in contact" when the inter-subunit distance between their Cβ-atoms is less than 12 Å. x ij and y ij are 1 if residue i and j are in contact in protein x (model) and y (native), respectively, and 0 otherwise. Σf(x ij , y ij ) represents the number of correctly predicted contacts, and Σg(x ij , y ij ) represents the number of unions of residue pairs in protein x and y. The MM-score is a multiple-chain extension of TM-score [24] and calculated according to the following equation: where L is the total length of all chains in the target and L ali is the number of aligned residue pairs between the native and the model structures. d ij is the distance between the Ca atoms of the aligned residues i and j after superposition of the native structure and the model. d 0 (L) is a parameter to correct the effect of chain length:

Targets for performance evaluation
Our methods were tested in the CASP9 experiment. For all 129 targets, 60 were human/server targets and 69 were server-only targets. From these, we excluded the following targets for our analyses: (1) structures with coordinates not released on April 7, 2011; (2) PDB structures with insertions in locations other than the N-and C-termini compared to the corresponding CASP target sequence; (3) structures with coordinates obtained by methods other than X-ray diffraction analysis; (4) targets determined as monomers from PDB headers of the native structure (described below); and (5) the models that were submitted as monomer by our group. Finally, we have 16 targets in human/server category and 58 targets in both human/server and server-only categories for our analyses.
It is notable that we submitted our models as monomer in cases when we predicted the target as monomer; we could not build reliable single-chain model for the target; CASP organizers gave us information that the structure of the target was solved by methods other than X-ray diffraction analysis; or we could not obtain enough information about the quaternary structure of the target from our template search.

Summary of the prediction results
Although the predictors in CASP9 could submit up to 5 models with ranking for each target, only the first model (i.e. the model the predictors believed to be the most probable) was considered in the official assessment. We mention only the first model in this paper. Table 1 shows the summary of our predictions provided by the auto server and by the manual method. Figure 1 shows examples of our models and the native structures. Note that human/server targets were more difficult than server-only targets because CASP organizers tended to avoid assigning easy targets for human prediction. With the auto server, we submitted our first models as oligomers for 58 targets in human/server and server-only categories. Out of them, 46 models had the correct number of chains and 14 models had correct quaternary structures. Here, we determined that the quaternary structure of a model is "correct" when S agree ≥ 0.25 and MM-score ≥ 0.5 according to the visual inspection, as described below. For prediction of human/server targets, we submitted our first models as oligomers for 25 targets. Of these, 19 models had the correct number of chains and 4 models had correct quaternary structures. The best server-only model (T0522) satisfies S agree ≥ 0.7 and MM-scores ≥ 0.8, showing that our auto server could generate rather accurate quaternary structure models under this completely blind test.
With the manual method, we submitted our first models as oligomers for 16 targets in human/server category. Out of them, 15 models had the correct number of chains and 11 models had correct quaternary structures. Among them, 2 models had S agree ≥ 0.7 and MM-score ≥ 0.8.
There were 15 multimer targets for which both our auto server and manual prediction submitted first models as multimers. To assess the effect of voting to determine the number of chains in our auto server predictions, the predicted number of chains for these 15 targets are listed in Table 2. Fourth column of Table 2 shows the number of chains predicted by the number of top hit templates of HHsearch (quaternary structures were determined by PISA) for comparison. Among these 15 targets, 8 were predicted correctly. On the other hand, the numbers of correctly predicted targets by our auto server and manual method were 11 and 14, respectively. These results show the effectiveness of our approaches. For these targets, the predicted number of chains, S agree , and MM-score of the models from auto server prediction and manual prediction are also listed in Table 3. Among the 15 targets, models from our manual prediction outperform those from the auto server for 11 targets with respect to S agree and 12 targets with respect to the MM-score. Figure 2 shows the plot for comparison of MM-score for the auto server and manual prediction per target. Figure 1 shows examples of the submitted models. T0542 is a server-only target (Figure 1(a)). The PDB ID of the native structure is 3n05. The protein is a 2-mer with 590 amino acids per chain. The top hit of the HHsearch fold recognition tool in our server was 3dla (8-mer). However, our server discarded this hit because the result of the prediction of the number of chains by voting Figure.1 Examples of the predictions made by our methods Red and blue indicate the predicted models, and gray indicates the native structures. Each predicted model is superimposed on the native structure for the entire structure except for (c) which indicates 1 subunit. (a) A successful example of the auto server (T0542, 3n05). (b) A successful example of the manual method (T0576, 3na2). (c) A failed example of the manual method (T0520, 3mr7). Although the tertiary structure of each monomer was well predicted, the spatial arrangement of the dimer was incorrect.  indicated that the most probable oligomeric state for this target was a 2-mer, and selected 3ilv (2-mer) as the template instead. This led to the correct prediction of both the monomer and the quaternary structure of this target, with an S agree of 0.298 and an MM-score of 0.856. This is an example that shows the framework of our server prediction works well. T0576 is a human/server target (Figure 1(b)). The PDB ID of the native structure is 3na2. The protein is a 2-mer with 172 amino acids per chain. The two chains interact with each other at their beta sheets and if quaternary structure is not formed, hydrophobic residues of these regions will be exposed. So quaternary structure is essential for this protein. The original sequence of this protein included a short tag sequence at the N-terminus and the effects of this tag led to false hits in sequence based template search methods, e.g. 2grg or 1z1s by HHsearch. Prediction by our auto server, which chose 1u83 as the template, also failed because of the same reason (S agree and the MM-score were 0.004 and 0.265, respectively). On the other hand, in our manual prediction, we eliminated the tag region from query of HHsearch. The top hit provided by this improved HHsearch was 3fm2 (2-mer) and the second hit was 2hqv (4-mer). We found that the corresponding residues of the interface of 2hqv were probably missing in this target according to the alignment of 2hqv and the target, so we selected 3fm2 to generate the quaternary structure of this target. S agree and the MM-score of our model have rather high values of 0.826 and 0.724, respectively. This shows an example of the effectiveness of manual inspection for template-based quaternary structure prediction.

Number
T0520 is an example of a failure of manual prediction. T0520 is a human/server target (Figure 1(c)). The PDB ID of the native structure is 3mr7. The protein is a 2-mer with 189 amino acids per chain. There are several templates, including 1wc3, 1ybt, and 2w01, which have structures similar to the first model of our auto server prediction, which we originally considered as the most probable monomer structure. This prediction was accurate. Among these templates, the spacial arrangement of chains 2w01 and 1wc3 are similar, so we modeled our quaternary structure based on these 2 templates. The native structure, however, has a chain orientation that is different from the chain orientation in all templates in spite of their similarity as monomers. The TM-score between the native structure and our model provided by human prediction was 0.810, but the S agree was 0.002 and the MM-score was 0.490.

Server prediction versus manual prediction
The ratios of the correctly predicted number of chains for human/ server targets were 19/25 (0.76) and 15/16 (0.94) for auto server prediction and manual prediction, respectively ( Table 3). The average of S agree for the auto server and manual prediction were 0.167 and 0.378, and that of the MM-scores were 0.492 and 0.648, respectively. In our predictions, manual prediction outperforms auto server prediction for all of these measures. This indicates that human intervention, i.e.

Figure.2 Auto server vs. manual method
Comparison of MM-score of the models by manual method and auto server prediction.   Figure.3 Scatter plot between Sagree and MM-score for the prediction targets in Table 3.
The closed circles are targets that we regarded as correct by visual inspection, and the open triangles are others. The correlation coefficient for all points was 0.787.

Table.3 -Comparison of prediction performance between auto server and manual method
more accurate chain number predictions, re-selection of templates, re-alignment between sequences of targets and templates, and avoiding collisions could considerably improve the prediction performance in template-based quaternary structure prediction, as described in the example of T0576 above.

S agree and MM-score
In this paper, we used 2 scores (S agree and MM-score; S agree was used in the official assessment of CASP9) to provide quality assessments of the predicted models. Since an assessment method for quaternary structure prediction has not been established, we discuss these 2 scoring measures. Figure 3 is a scatter plot between S agree and MM-score for the data in Table 3. These were moderately correlated with each other, and the correlation coefficient was 0.787. The closed circles in Figure 3 are the targets judged by visual inspection to have correctly predicted quaternary structures. According to this judgment, despite some exceptions, predicted models with an S agree value over 0.25 and an MM-score over 0.5 could be considered as good models. Note that a score called IS-score [26] that can consider significance of similarity between target and model structures compared to random pairs was developed recently by the same group as MM-score. They reported that all near native docking models in CAPRI have an IS-score above 0.17. According to this threshold, two models (T0517 and T0592) among our "good" models in Table 3 have IS-scores below 0.17 and two models (T0602, T0629) among our "not good" models in Table 3 have IS-scores above 0.17. However, classification of "good" and "not good" for other targets was the same, and we found that IS-score is correlated with S agree and also MM-score (correlation coefficients are 0.89 between S agree and IS-score and 0.73 between MM-score and IS-score; data not shown). This also supports that our criteria of "good models" are appropriate. There are 3 types of prediction failures: (1) incorrect number of chains, (2) incorrect arrangement of subunits, and (3) low quality of each monomer of the model. (1) When the number of chains is incorrect, S agree is extremely low while the MM-score appears to be reasonably low. This can be seen by comparing predictions between the auto server and the manual method for T0517 and T0592.
(2) When the spatial arrangement of subunits is incorrect, S agree is extremely low but the MM-score is not significantly low, as we can see for T0520 described above and also for T0596. To improve the prediction for these targets, a detailed assessment of the interfaces in a model is needed.
(3) When a portion of each monomer is not well modeled and the portion is located near the subunit interface, both S agree and MMscore tend to be low. Predicted models of T0517, T0602, and T0629 by the manual method are examples of this scenario.

Conclusions
We have discussed the overall performance of our local tools, successful examples as well as unsuccessful examples, and current challenges. Our quaternary structure predictions in CASP9 shows doi: 10.7243/2050-2273-1-1 the effectiveness of the template-based method in this field. As far as we know, this is the first systematic performance assessment of completely blind protein quaternary structure prediction from its amino acid sequence. We have demonstrated that simple auto server prediction could generate successful models for 14 out of 58 targets. Human experts could generate better models than our auto server through more accurate chain number predictions, reselection of templates, re-alignment between sequences of targets and templates, and avoiding collisions. We have also shown in several cases that good monomer templates could provide different chain orientations, leading to poor quality of predicted quaternary structures based on templates. Our template-based quaternary structure prediction methods are simple and general and therefore can be expanded further with various options or combinations with other methods. Our results should be useful for improvement of the accuracy of template-based quaternary structure prediction.
Authors' contributions MM, MK, and SN have developed the protein tertiary and quaternary structure prediction server. MM and MK performed manual prediction of protein quaternary structures, and SN carried out result analysis. KS helped to draft the manuscript. All authors read and approved the final manuscript.