Classifier ensembles for protein structural class prediction with varying homology

Structural class characterizes the overall folding type of a protein or its domain. A number of computational methods have been proposed to predict structural class based on primary sequences; however, the accuracy of these methods is strongly affected by sequence homology. This paper proposes, an ensemble classification method and a compact feature-based sequence representation. This method improves prediction accuracy for the four main structural classes compared to competing methods, and provides highly accurate predictions for sequences of widely varying homologies. The experimental evaluation of the proposed method shows superior results across sequences that are characterized by entire homology spectrum, ranging from 25% to 90% homology. The error rates were reduced by over 20% when compared with using individual prediction methods and most commonly used composition vector representation of protein sequences. Comparisons with competing methods on three large benchmark datasets consistently show the superiority of the proposed method.     

Accurate prediction of protein structural classes using functional domains and predicted secondary structure sequences

Protein structural class prediction is one of the challenging problems in bioinformatics. Previous methods directly based on the similarity of amino acid (AA) sequences have been shown to be insufficient for low-similarity protein data-sets. To improve the prediction accuracy for such low-similarity proteins, different methods have been recently proposed that explore the novel feature sets based on predicted secondary structure propensities. In this paper, we focus on protein structural class prediction using combinations of the novel features including secondary structure propensities as well as functional domain (FD) features extracted from the InterPro signature database. Our comprehensive experimental results based on several benchmark data-sets have shown that the integration of new FD features substantially improves the accuracy of structural class prediction for low-similarity proteins as they capture meaningful relationships among AA residues that are far away in protein sequence. The proposed prediction method has also been tested to predict structural classes for partially disordered proteins with the reasonable prediction accuracy, which is a more difficult problem comparing to structural class prediction for commonly used benchmark data-sets and has never been done before to the best of our knowledge. In addition, to avoid overfitting with a large number of features, feature selection is applied to select discriminating features that contribute to achieve high prediction accuracy. The selected features have been shown to achieve stable prediction performance across different benchmark data-sets.     

Identification of related proteins on family, superfamily and fold level

Proteins might have considerable structural similarities even when no evolutionary relationship of their sequences can be detected. This property is often referred to as the proteins sharing only a "fold". Of course, there are also sequences of common origin in each fold, called a "superfamily", and in them groups of sequences with clear similarities, designated "family". Developing algorithms to reliably identify proteins related at any level is one of the most important challenges in the fast growing field of bioinformatics today. However, it is not at all certain that a method proficient at finding sequence similarities performs well at the other levels, or vice versa.Here, we have compared the performance of various search methods on these different levels of similarity. As expected, we show that it becomes much harder to detect proteins as their sequences diverge. For family related sequences the best method gets 75% of the top hits correct. When the sequences differ but the proteins belong to the same superfamily this drops to 29%, and in the case of proteins with only fold similarity it is as low as 15%. We have made a more complete analysis of the performance of different algorithms than earlier studies, also including threading methods in the comparison. Using this method a more detailed picture emerges, showing multiple sequence information to improve detection on the two closer levels of relationship. We have also compared the different methods of including this information in prediction algorithms.For lower specificities, the best scheme to use is a linking method connecting proteins through an intermediate hit. For higher specificities, better performance is obtained by PSI-BLAST and some procedures using hidden Markov models. We also show that a threading method, THREADER, performs significantly better than any other method at fold recognition.     

Cross-validation of protein structural class prediction using statistical clustering and neural networks.

We present an approach to predicting protein structural class that uses amino acid composition and hydrophobic pattern frequency information as input to two types of neural networks: (1) a three-layer back-propagation network and (2) a learning vector quantization network. The results of these methods are compared to those obtained from a modified Euclidean statistical clustering algorithm. The protein sequence data used to drive these algorithms consist of the normalized frequency of up to 20 amino acid types and six hydrophobic amino acid patterns. From these frequency values the structural class predictions for each protein (all-alpha, all-beta, or alpha-beta classes) are derived. Examples consisting of 64 previously classified proteins were randomly divided into multiple training (56 proteins) and test (8 proteins) sets. The best performing algorithm on the test sets was the learning vector quantization network using 17 inputs, obtaining a prediction accuracy of 80.2%. The Matthews correlation coefficients are statistically significant for all algorithms and all structural classes. The differences between algorithms are in general not statistically significant. These results show that information exists in protein primary sequences that is easily obtainable and useful for the prediction of protein structural class by neural networks as well as by standard statistical clustering algorithms.     

Prediction of protein structural class for the twilight zone sequences

Structural class characterizes the overall folding type of a protein or its domain. This paper develops an accurate method for in silico prediction of structural classes from low homology (twilight zone) protein sequences. The proposed LLSC-PRED method applies linear logistic regression classifier and a custom-designed, feature-based sequence representation to provide predictions. The main advantages of the LLSC-PRED are the comprehensive representation that includes 58 features describing composition and physicochemical properties of the sequences and transparency of the prediction model. The representation also includes predicted secondary structure content, thus for the first time exploring synergy between these two related predictions. Based on tests performed with a large set of 1673 twilight zone domains, the LLSC-PRED's prediction accuracy, which equals over 62%, is shown to be better than accuracy of over a dozen recently published competing in silico methods and similar to accuracy of other, non-transparent classifiers that use the proposed representation.     

A Consensus Data Mining secondary structure prediction by combining GOR V and Fragment Database Mining

The major aim of tertiary structure prediction is to obtain protein models with the highest possible accuracy. Fold recognition, homology modeling, and de novo prediction methods typically use predicted secondary structures as input, and all of these methods may significantly benefit from more accurate secondary structure predictions. Although there are many different secondary structure prediction methods available in the literature, their cross-validated prediction accuracy is generally <80%. In order to increase the prediction accuracy, we developed a novel hybrid algorithm called Consensus Data Mining (CDM) that combines our two previous successful methods: (1) Fragment Database Mining (FDM), which exploits the Protein Data Bank structures, and (2) GOR V, which is based on information theory, Bayesian statistics, and multiple sequence alignments (MSA). In CDM, the target sequence is dissected into smaller fragments that are compared with fragments obtained from related sequences in the PDB. For fragments with a sequence identity above a certain sequence identity threshold, the FDM method is applied for the prediction. The remainder of the fragments are predicted by GOR V. The results of the CDM are provided as a function of the upper sequence identities of aligned fragments and the sequence identity threshold. We observe that the value 50% is the optimum sequence identity threshold, and that the accuracy of the CDM method measured by Q(3) ranges from 67.5% to 93.2%, depending on the availability of known structural fragments with sufficiently high sequence identity. As the Protein Data Bank grows, it is anticipated that this consensus method will improve because it will rely more upon the structural fragments.     

Exploring an alignment free approach for protein classification and structural class prediction

Alignment free methods based on Chaos Game Representation (CGR), also known as sequence signature approaches, have proven of great interest for DNA sequence analysis. Indeed, they have been successfully applied for sequence comparison, phylogeny, detection of horizontal transfers or extraction of representative motifs in regulation sequences. Transposing such methods to proteins poses several fundamental questions related to representation space dimensionality. Several studies have tackled these points, but none has, so far, brought the application of CGRs to proteins to their fully expected potential. Yet, several studies have shown that techniques based on n-peptide frequencies can be relevant for proteins. Here, we investigate the effectiveness of a strategy based on the CGR approach using a fixed reverse encoding of amino acids into nucleic sequences. We first explore its relevance to protein classification into functional families. We then attempt to apply it to the prediction of protein structural classes. Our results suggest that the reverse encoding approach could be relevant in both cases. We show that it is able to classify functional families of proteins by extracting signatures close to the ProSite patterns. Applied to structural classification, the approach reaches scores of correct classification close to 84%, i.e. close to the scores of related methods in the field. Various optimizations of the approach are still possible, which open the door for future applications.     

SVM-based method for protein structural class prediction using secondary structural content and structural information of amino acids

The knowledge collated from the known protein structures has revealed that the proteins are usually folded into the four structural classes: all-α, all-β, α/β and α + β. A number of methods have been proposed to predict the protein's structural class from its primary structure; however, it has been observed that these methods fail or perform poorly in the cases of distantly related sequences. In this paper, we propose a new method for protein structural class prediction using low homology (twilight-zone) protein sequences dataset. Since protein structural class prediction is a typical classification problem, we have developed a Support Vector Machine (SVM)-based method for protein structural class prediction that uses features derived from the predicted secondary structure and predicted burial information of amino acid residues. The examination of different individual as well as feature combinations revealed that the combination of secondary structural content, secondary structural and solvent accessibility state frequencies of amino acids gave rise to the best leave-one-out cross-validation accuracy of ~81% which is comparable to the best accuracy reported in the literature so far.     

Predicting protein structural class by SVM with class-wise optimized features and decision probabilities

Determination of protein structural class solely from sequence information is a challenging task. Several attempts to solve this problem using various methods can be found in literature. We present support vector machine (SVM) approach where probability-based decision is used along with class-wise optimized feature sets. This approach has two distinguishing characteristics from earlier attempts: (1) it uses class-wise optimized features and (2) decisions of different SVM classifiers are coupled with probability estimates to make the final prediction. The algorithm was tested on three datasets, containing 498 domains, 1092 domains and 5261 domains. Ten-fold external cross-validation was performed to assess the performance of the algorithm. Significantly high accuracy of 92.89% was obtained for the 498-dataset. We achieved 54.67% accuracy for the dataset with 1092 domains, which is better than the previously reported best accuracy of 53.8%. We obtained 59.43% prediction accuracy for the larger and less redundant 5261-dataset. We also investigated the advantage of using class-wise features over union of these features (conventional approach) in one-vs.-all SVM framework. Our results clearly show the advantage of using class-wise optimized features. Brief analysis of the selected class-wise features indicates their biological significance.     

Highly specific localization of promoter regions in large genomic sequences by PromoterInspector: a novel context analysis approach

We present a new algorithm called PromoterInspector to locate eukaryotic polymase II promoter regions in large genomic sequences with a high degree of specificity. PromoterInspector focuses on the genetic context of promoters, rather than their exact location. Application of PromoterInspector can serve as a crucial pre-processing step for other methods to locate exactly, or to analyze promoters.PromoterInspector does not depend on heuristics, because it is purely based on libraries of IUPAC words extracted from training sequences by an unsupervised learning approach. We compared PromoterInspector to in silico promoter prediction tools using the sequences from the review by J.W. Fickett. PromoterInspector compared favourably on Fickett's evaluation scheme. A true positive to false positive ratio of 2.3 was obtained, surpassing the best ratio of 0.6, reported for TSSG. The application of our method to several large genomic sequences of over 1.3 million base-pairs in total resulted in even more specific predictions. The coverage of annotated promoters was comparable to other in silico promoter prediction methods, while the true positive predictions increased by up to 100% of total matches.PromoterInspector scans 100 kb in less than one minute on a workstation, and thus is especially applicable for large genome analysis. The method is available at http://genomatix.gsf. de/cgi-bin/promoterinspector/promoterinspector.pl.     

A novel approach to fold recognition using sequence-derived properties from sets of structurally similar local fragments of proteins

Comparative modeling methods can consistently produce reliable structural models for protein sequences with more than 25% sequence identity to proteins with known structure. However, there is a good chance that also sequences with lower sequence identity have their structural components represented in structural databases. To this end, we present a novel fragment-based method using sets of structurally similar local fragments of proteins. The approach differs from other fragment-based methods that use only single backbone fragments. Instead, we use a library of groups containing sets of sequence fragments with geometrically similar local structures and extract sequence related properties to assign these specific geometrical conformations to target sequences. We test the ability of the approach to recognize correct SCOP folds for 273 sequences from the 49 most popular folds. 49% of these sequences have the correct fold as their top prediction, while 82% have the correct fold in one of the top five predictions. Moreover, the approach shows no performance reduction on a subset of sequence targets with less than 10% sequence identity to any protein used to build the library.     

Origins of structural diversity within sequentially identical hexapeptides

Efforts to predict protein secondary structure have been hampered by the apparent structural plasticity of local amino acid sequences. Kabsch and Sander (1984, Proc. Natl. Acad. Sci. USA 81, 1075–1078) articulated this problem by demonstrating that identical pentapeptide sequences can adopt distinct structures in different proteins. With the increased size of the protein structure database and the availability of new methods to characterize structural environments, we revisit this observation of structural plasticity. Within a set of proteins with less than 50% sequence identity, 59 pairs of identical hexapeptide sequences were identified. These local structures were compared and their surrounding structural environments examined. Within a protein structural class (α/α, β/β, α/β, α + β), the structural similarity of sequentially identical hexapeptides usually is preserved. This study finds eight pairs of identical hexapeptide sequences that adopt β-strand structure in one protein and α-helical structure in the other. In none of the eight cases do the members of these sequence pairs come from proteins within the same folding class. These results have implications for class dependent secondary structure prediction algorithms.     

Identifying complete RNA structural ensembles including pseudoknots

The close relationship between RNA structure and function underlines the significance of accurately predicting RNA structures from sequence information. Structural topologies such as pseudoknots are of particular interest due to their ubiquity and direct involvement in RNA function, but identifying pseudoknots is a computationally challenging problem and existing heuristic approaches usually perform poorly for RNA sequences of even a few hundred bases. We survey the performance of pseudoknot prediction methods on a data set of full-length RNA sequences representing varied sequence lengths, and biological RNA classes such as RNase P RNA, Group I Intron, tmRNA and tRNA. Pseudoknot prediction methods are compared with minimum free energy and suboptimal secondary structure prediction methods in terms of correct base-pairs, stems and pseudoknots and we find that the ensemble of suboptimal structure predictions succeeds in identifying correct structural elements in RNA that are usually missed in MFE and pseudoknot predictions. We propose a strategy to identify a comprehensive set of non-redundant stems in the suboptimal structure space of a RNA molecule by applying heuristics that reduce the structural redundancy of the predicted suboptimal structures by merging slightly varying stems that are predicted to form in local sequence regions. This reduced-redundancy set of structural elements consistently outperforms more specialized approaches.in data sets. Thus, the suboptimal folding space can be used to represent the structural diversity of an RNA molecule more comprehensively than optimal structure prediction approaches alone.     

Retrovirus integrase: identification of a potential leucine zipper motif.

The secondary structure of the retrovirus integration protein (IN) was predicted from seven inferred retrovirus IN sequences. The IN sequences were aligned by computer and the phylogenetic relationships between them were determined. The secondary structure of the aligned IN sequences was predicted by two consensus prediction methods. The predicted secondary structural patterns from the two consensus prediction schemes were compared with and superimposed on a composite structural profile of hydropathic/chain flexibility/amphipathic indexes with each index profile being calculated independently for the aligned IN sequences. The use of this composite structural profile not only enhanced the prediction accuracy but also helped in defining the surface loop regions which would be otherwise unpredictable by the use of consensus prediction methods alone. An amphipathic helix was identified by these united structural prediction-chain property profiles. Helical wheel analysis gave the amphipathic helix a coiled-coil like pattern which was similar to the leucine zipper discovered for some eukaryotic gene regulatory proteins. The proposed amphipathic helix may play an essential role in defining the biological properties of IN.     

Rich parameterization improves RNA structure prediction

Current approaches to RNA structure prediction range from physics-based methods, which rely on thousands of experimentally measured thermodynamic parameters, to machine-learning (ML) techniques. While the methods for parameter estimation are successfully shifting toward ML-based approaches, the model parameterizations so far remained fairly constant. We study the potential contribution of increasing the amount of information utilized by RNA folding prediction models to the improvement of their prediction quality. This is achieved by proposing novel models, which refine previous ones by examining more types of structural elements, and larger sequential contexts for these elements. Our proposed fine-grained models are made practical thanks to the availability of large training sets, advances in machine-learning, and recent accelerations to RNA folding algorithms. We show that the application of more detailed models indeed improves prediction quality, while the corresponding running time of the folding algorithm remains fast. An additional important outcome of this experiment is a new RNA folding prediction model (coupled with a freely available implementation), which results in a significantly higher prediction quality than that of previous models. This final model has about 70,000 free parameters, several orders of magnitude more than previous models. Being trained and tested over the same comprehensive data sets, our model achieves a score of 84% according to the F?-measure over correctly-predicted base-pairs (i.e., 16% error rate), compared to the previously best reported score of 70% (i.e., 30% error rate). That is, the new model yields an error reduction of about 50%. Trained models and source code are available at www.cs.bgu.ac.il/?negevcb/contextfold.     

Accurate prediction of protein structural classes using functional domains and predicted secondary structure sequences

Protein structural class prediction is one of the challenging problems in bioinformatics. Previous methods directly based on the similarity of amino acid (AA) sequences have been shown to be insufficient for low-similarity protein data-sets. To improve the prediction accuracy for such low-similarity proteins, different methods have been recently proposed that explore the novel feature sets based on predicted secondary structure propensities. In this paper, we focus on protein structural class prediction using combinations of the novel features including secondary structure propensities as well as functional domain (FD) features extracted from the InterPro signature database. Our comprehensive experimental results based on several benchmark data-sets have shown that the integration of new FD features substantially improves the accuracy of structural class prediction for low-similarity proteins as they capture meaningful relationships among AA residues that are far away in protein sequence. The proposed prediction method has also been tested to predict structural classes for partially disordered proteins with the reasonable prediction accuracy, which is a more difficult problem comparing to structural class prediction for commonly used benchmark data-sets and has never been done before to the best of our knowledge. In addition, to avoid overfitting with a large number of features, feature selection is applied to select discriminating features that contribute to achieve high prediction accuracy. The selected features have been shown to achieve stable prediction performance across different benchmark data-sets.     

Prediction of protein structural class for low-similarity sequences using support vector machine and PSI-BLAST profile

Knowledge of structural class plays an important role in understanding protein folding patterns. In this study, a simple and powerful computational method, which combines support vector machine with PSI-BLAST profile, is proposed to predict protein structural class for low-similarity sequences. The evolution information encoding in the PSI-BLAST profiles is converted into a series of fixed-length feature vectors by extracting amino acid composition and dipeptide composition from the profiles. The resulting vectors are then fed to a support vector machine classifier for the prediction of protein structural class. To evaluate the performance of the proposed method, jackknife cross-validation tests are performed on two widely used benchmark datasets, 1189 (containing 1092 proteins) and 25PDB (containing 1673 proteins) with sequence similarity lower than 40% and 25%, respectively. The overall accuracies attain 70.7% and 72.9% for 1189 and 25PDB datasets, respectively. Comparison of our results with other methods shows that our method is very promising to predict protein structural class particularly for low-similarity datasets and may at least play an important complementary role to existing methods.     

Prediction of β-turns and β-turn types by a novel bidirectional Elman-type recurrent neural network with multiple output layers (MOLEBRNN)

Prediction of β-turns from amino acid sequences has long been recognized as an important problem in structural bioinformatics due to their frequent occurrence as well as their structural and functional significance. Because various structural features of proteins are intercorrelated, secondary structure information has been often employed as an additional input for machine learning algorithms while predicting β-turns. Here we present a novel bidirectional Elman-type recurrent neural network with multiple output layers (MOLEBRNN) capable of predicting multiple mutually dependent structural motifs and demonstrate its efficiency in recognizing three aspects of protein structure: β-turns, β-turn types, and secondary structure. The advantage of our method compared to other predictors is that it does not require any external input except for sequence profiles because interdependencies between different structural features are taken into account implicitly during the learning process. In a sevenfold cross-validation experiment on a standard test dataset our method exhibits the total prediction accuracy of 77.9% and the Mathew's Correlation Coefficient of 0.45, the highest performance reported so far. It also outperforms other known methods in delineating individual turn types. We demonstrate how simultaneous prediction of multiple targets influences prediction performance on single targets. The MOLEBRNN presented here is a generic method applicable in a variety of research fields where multiple mutually depending target classes need to be predicted. Availability: http://webclu.bio.wzw.tum.de/predator-web/.     

Customised fragments libraries for protein structure prediction based on structural class annotations

Background

Since experimental techniques are time and cost consuming, in silico protein structure prediction is essential to produce conformations of protein targets. When homologous structures are not available, fragment-based protein structure prediction has become the approach of choice. However, it still has many issues including poor performance when targets’ lengths are above 100 residues, excessive running times and sub-optimal energy functions. Taking advantage of the reliable performance of structural class prediction software, we propose to address some of the limitations of fragment-based methods by integrating structural constraints in their fragment selection process.

Results

Using Rosetta, a state-of-the-art fragment-based protein structure prediction package, we evaluated our proposed pipeline on 70 former CASP targets containing up to 150 amino acids. Using either CATH or SCOP-based structural class annotations, enhancement of structure prediction performance is highly significant in terms of both GDT_TS (at least +2.6, p-values < 0.0005) and RMSD (−0.4, p-values < 0.005). Although CATH and SCOP classifications are different, they perform similarly. Moreover, proteins from all structural classes benefit from the proposed methodology. Further analysis also shows that methods relying on class-based fragments produce conformations which are more relevant to user and converge quicker towards the best model as estimated by GDT_TS (up to 10% in average). This substantiates our hypothesis that usage of structurally relevant templates conducts to not only reducing the size of the conformation space to be explored, but also focusing on a more relevant area.

Conclusions

Since our methodology produces models the quality of which is up to 7% higher in average than those generated by a standard fragment-based predictor, we believe it should be considered before conducting any fragment-based protein structure prediction. Despite such progress, ab initio prediction remains a challenging task, especially for proteins of average and large sizes. Apart from improving search strategies and energy functions, integration of additional constraints seems a promising route, especially if they can be accurately predicted from sequence alone.

Electronic supplementary material

The online version of this article (doi:10.1186/s12859-015-0576-2) contains supplementary material, which is available to authorized users.     

基于蛋白质二级结构内容的域结构类预测

运用加入竞争层的BP网络,研究了基于蛋白质二级结构内容的域结构类预测问题.在BP网络中嵌入一竞争,层显著提高了网络预测性能.仅使用了一个小的训练集和简单的网络结构,获得了很高的预测精度自支持精度97.62%,jack-knife测试精度97.62%,及平均外推精度90.74%.在建立更完备的域结构类特征向量和更有代表性的训练集的基础上,所述方法将为蛋白质域结构分类领域提供新的分类基准.