Prediction of contact maps with neural networks and correlated mutations.

Contact maps of proteins are predicted with neural network-based methods, using as input codings of increasing complexity including evolutionary information, sequence conservation, correlated mutations and predicted secondary structures. Neural networks are trained on a data set comprising the contact maps of 173 non-homologous proteins as computed from their well resolved three-dimensional structures. Proteins are selected from the Protein Data Bank database provided that they align with at least 15 similar sequences in the corresponding families. The predictors are trained to learn the association rules between the covalent structure of each protein and its contact map with a standard back propagation algorithm and tested on the same protein set with a cross-validation procedure. Our results indicate that the method can assign protein contacts with an average accuracy of 0.21 and with an improvement over a random predictor of a factor >6, which is higher than that previously obtained with methods only based either on neural networks or on correlated mutations. Furthermore, filtering the network outputs with a procedure based on the residue coordination numbers, the accuracy of predictions increases up to 0.25 for all the proteins, with an 8-fold deviation from a random predictor. These scores are the highest reported so far for predicting protein contact maps.     

Prediction of protein interaction sites from sequence profile and residue neighbor list.

Protein-protein interaction sites are predicted from a neural network with sequence profiles of neighboring residues and solvent exposure as input. The network was trained on 615 pairs of nonhomologous complex-forming proteins. Tested on a different set of 129 pairs of nonhomologous complex-forming proteins, 70% of the 11,004 predicted interface residues are actually located in the interfaces. These 7732 correctly predicted residues account for 65% of the 11,805 residues making up the 129 interfaces. The main strength of the network predictor lies in the fact that neighbor lists and solvent exposure are relatively insensitive to structural changes accompanying complex formation. As such, it performs equally well with bound or unbound structures of the proteins. For a set of 35 test proteins, when the input was calculated from the bound and unbound structures, the correct fractions of the predicted interface residues were 69 and 70%, respectively.     

Prediction of interface residues in protein-protein complexes by a consensus neural network method: test against NMR data

The number of structures of protein-protein complexes deposited to the Protein Data Bank is growing rapidly. These structures embed important information for predicting structures of new protein complexes. This motivated us to develop the PPISP method for predicting interface residues in protein-protein complexes. In PPISP, sequence profiles and solvent accessibility of spatially neighboring surface residues were used as input to a neural network. The network was trained on native interface residues collected from the Protein Data Bank. The prediction accuracy at the time was 70% with 47% coverage of native interface residues. Now we have extensively improved PPISP. The training set now consisted of 1156 nonhomologous protein chains. Test on a set of 100 nonhomologous protein chains showed that the prediction accuracy is now increased to 80% with 51% coverage. To solve the problem of over-prediction and under-prediction associated with individual neural network models, we developed a consensus method that combines predictions from multiple models with different levels of accuracy and coverage. Applied on a benchmark set of 68 proteins for protein-protein docking, the consensus approach outperformed the best individual models by 3-8 percentage points in accuracy. To demonstrate the predictive power of cons-PPISP, eight complex-forming proteins with interfaces characterized by NMR were tested. These proteins are nonhomologous to the training set and have a total of 144 interface residues identified by chemical shift perturbation. cons-PPISP predicted 174 interface residues with 69% accuracy and 47% coverage and promises to complement experimental techniques in characterizing protein-protein interfaces. .     

DISPLAR: an accurate method for predicting DNA-binding sites on protein surfaces

Structural and physical properties of DNA provide important constraints on the binding sites formed on surfaces of DNA-targeting proteins. Characteristics of such binding sites may form the basis for predicting DNA-binding sites from the structures of proteins alone. Such an approach has been successfully developed for predicting protein–protein interface. Here this approach is adapted for predicting DNA-binding sites. We used a representative set of 264 protein–DNA complexes from the Protein Data Bank to analyze characteristics and to train and test a neural network predictor of DNA-binding sites. The input to the predictor consisted of PSI-blast sequence profiles and solvent accessibilities of each surface residue and 14 of its closest neighboring residues. Predicted DNA-contacting residues cover 60% of actual DNA-contacting residues and have an accuracy of 76%. This method significantly outperforms previous attempts of DNA-binding site predictions. Its application to the prion protein yielded a DNA-binding site that is consistent with recent NMR chemical shift perturbation data, suggesting that it can complement experimental techniques in characterizing protein–DNA interfaces.     

Prediction of distant residue contacts with the use of evolutionary information

In this work we present a novel correlated mutations analysis (CMA) method that is significantly more accurate than previously reported CMA methods. Calculation of correlation coefficients is based on physicochemical properties of residues (predictors) and not on substitution matrices. This results in reliable prediction of pairs of residues that are distant in protein sequence but proximal in its three dimensional tertiary structure. Multiple sequence alignments (MSA) containing a sequence of known structure for 127 families from PFAM database have been selected so that all major protein architectures described in CATH classification database are represented. Protein sequences in the selected families were filtered so that only those evolutionarily close to the target protein remain in the MSA. The average accuracy obtained for the alpha beta class of proteins was 26.8% of predicted proximal pairs with average improvement over random accuracy (IOR) of 6.41. Average accuracy is 20.6% for the mainly beta class and 14.4% for the mainly alpha class. The optimum correlation coefficient cutoff (cc cutoff) was found to be around 0.65. The first predictor, which correlates to hydrophobicity, provides the most reliable results. The other two predictors give good predictions which can be used in conjunction to those of the first one. When stricter cc cutoff is chosen, the average accuracy increases significantly (38.76% for alpha beta class), but the trade off is a smaller number of predictions. The use of solvent accessible area estimations for filtering false positives out of the predictions is promising.     

Estimating the length of transmembrane helices using Z-coordinate predictions

Zpred2 is an improved version of ZPRED, a predictor for the Z-coordinates of alpha-helical membrane proteins, that is, the distance of the residues from the center of the membrane. Using principal component analysis and a set of neural networks, Zpred2 analyzes data extracted from the amino acid sequence, the predicted topology, and evolutionary profiles. Zpred2 achieves an average accuracy error of 2.18 A (2.17 A when an independent test set is used), an improvement by 15% compared to the previous version. We show that this accuracy is sufficient to enable the predictions of helix lengths with a correlation coefficient of 0.41. As a comparison, two state-of-the-art HMM-based topology prediction methods manage to predict the helix lengths with a correlation coefficient of less than 0.1. In addition, we applied Zpred2 to two other problems, the re-entrant region identification and model validation. Re-entrants were able to be detected with a certain consistency, but not better than with previous approaches, while incorrect models as well as mispredicted helices of transmembrane proteins could be distinguished based on the Z-coordinate predictions.     

Towards accurate residue-residue hydrophobic contact prediction for alpha helical proteins via integer linear optimization

A new optimization-based method is presented to predict the hydrophobic residue contacts in alpha-helical proteins. The proposed approach uses a high resolution distance dependent force field to calculate the interaction energy between different residues of a protein. The formulation predicts the hydrophobic contacts by minimizing the sum of these contact energies. These residue contacts are highly useful in narrowing down the conformational space searched by protein structure prediction algorithms. The proposed algorithm also offers the algorithmic advantage of producing a rank ordered list of the best contact sets. This model was tested on four independent alpha-helical protein test sets and was found to perform very well. The average accuracy of the predictions (separated by at least six residues) obtained using the presented method was approximately 66% for single domain proteins. The average true positive and false positive distances were also calculated for each protein test set and they are 8.87 and 14.67 A, respectively.     

Prediction of beta-turns in proteins from multiple alignment using neural network

A neural network-based method has been developed for the prediction of beta-turns in proteins by using multiple sequence alignment. Two feed-forward back-propagation networks with a single hidden layer are used where the first-sequence structure network is trained with the multiple sequence alignment in the form of PSI-BLAST-generated position-specific scoring matrices. The initial predictions from the first network and PSIPRED-predicted secondary structure are used as input to the second structure-structure network to refine the predictions obtained from the first net. A significant improvement in prediction accuracy has been achieved by using evolutionary information contained in the multiple sequence alignment. The final network yields an overall prediction accuracy of 75.5% when tested by sevenfold cross-validation on a set of 426 nonhomologous protein chains. The corresponding Q(pred), Q(obs), and Matthews correlation coefficient values are 49.8%, 72.3%, and 0.43, respectively, and are the best among all the previously published beta-turn prediction methods. The Web server BetaTPred2 (http://www.imtech.res.in/raghava/betatpred2/) has been developed based on this approach.     

i‐Patch: Interprotein contact prediction using local network information

Biological processes are commonly controlled by precise protein‐protein interactions. These connections rely on specific amino acids at the binding interfaces. Here we predict the binding residues of such interprotein complexes. We have developed a suite of methods, i‐Patch, which predict the interprotein contact sites by considering the two proteins as a network, with residues as nodes and contacts as edges. i‐Patch starts with two proteins, A and B, which are assumed to interact, but for which the structure of the complex is not available. However, we assume that for each protein, we have a reference structure and a multiple sequence alignment of homologues. i‐Patch then uses the propensities of patches of residues to interact, to predict interprotein contact sites. i‐Patch outperforms several other tested algorithms for prediction of interprotein contact sites. It gives 59% precision with 20% recall on a blind test set of 31 protein pairs. Combining the i‐Patch scores with an existing correlated mutation algorithm, McBASC, using a logistic model gave little improvement. Results from a case study, on bacterial chemotaxis protein complexes, demonstrate that our predictions can identify contact residues, as well as suggesting unknown interfaces in multiprotein complexes. Proteins 2010. © 2010 Wiley‐Liss, Inc.     

Prediction of helix-helix contacts and interacting helices in polytopic membrane proteins using neural networks

Despite rapidly increasing numbers of available 3D structures, membrane proteins still account for less than 1% of all structures in the Protein Data Bank. Recent high-resolution structures indicate a clearly broader structural diversity of membrane proteins than initially anticipated, motivating the development of reliable structure prediction methods specifically tailored for this class of molecules. One important prediction target capturing all major aspects of a protein's 3D structure is its contact map. Our analysis shows that computational methods trained to predict residue contacts in globular proteins perform poorly when applied to membrane proteins. We have recently published a method to identify interacting alpha-helices in membrane proteins based on the analysis of coevolving residues in predicted transmembrane regions. Here, we present a substantially improved algorithm for the same problem, which uses a newly developed neural network approach to predict helix-helix contacts. In addition to the input features commonly used for contact prediction of soluble proteins, such as windowed residue profiles and residue distance in the sequence, our network also incorporates features that apply to membrane proteins only, such as residue position within the transmembrane segment and its orientation toward the lipophilic environment. The obtained neural network can predict contacts between residues in transmembrane segments with nearly 26% accuracy. It is therefore the first published contact predictor developed specifically for membrane proteins performing with equal accuracy to state-of-the-art contact predictors available for soluble proteins. The predicted helix-helix contacts were employed in a second step to identify interacting helices. For our dataset consisting of 62 membrane proteins of solved structure, we gained an accuracy of 78.1%. Because the reliable prediction of helix interaction patterns is an important step in the classification and prediction of membrane protein folds, our method will be a helpful tool in compiling a structural census of membrane proteins.     

氨基酸邻域对二级结构的影响程度研究

通过研究神经网络权值矩阵的算法,挖掘蛋白质二级结构与氨基酸序列间的内在规律,提高一级序列预测二级结构的准确度。神经网络方法在特征分类方面具有良好表现,经过学习训练后的神经元连接权值矩阵包含样本的内在特征和规律。研究使用神经网络权值矩阵打分预测;采用错位比对方法寻找敏感的氨基酸邻域;分析测试集在不同加窗长度下的共性表现。实验表明,在滑动窗口长度L=7时,预测性能变化显著;邻域位置P=4的氨基酸残基对预测性能有加强作用。该研究方法为基于局部序列特征的蛋白质二级结构预测提供了新的算法设计。     

A Study on Protein Residue Contacts Prediction by Recurrent Neural Network

1IntroductionThe three-dimensional(3D)structure of a proteinis perhaps the most important of all its features,since itdetermines completely how the protein functions andinteracts with other molecules.Most biological mech-anisms at the protein level are based on shape-complementarity,so that proteins present particularconcavities and convexities that allow them to bind toeach other and formcomplexstructures,and tendon.Forthis reason,for instance,the drug design problem con-sists primarily in th…     

New methods for accurate prediction of protein secondary structure.

A primary and a secondary neural network are applied to secondary structure and structural class prediction for a database of 681 non-homologous protein chains. A new method of decoding the outputs of the secondary structure prediction network is used to produce an estimate of the probability of finding each type of secondary structure at every position in the sequence. In addition to providing a reliable estimate of the accuracy of the predictions, this method gives a more accurate Q3 (74.6%) than the cutoff method which is commonly used. Use of these predictions in jury methods improves the Q3 to 74.8%, the best available at present. On a database of 126 proteins commonly used for comparison of prediction methods, the jury predictions are 76.6% accurate. An estimate of the overall Q3 for a given sequence is made by averaging the estimated accuracy of the prediction over all residues in the sequence. As an example, the analysis is applied to the target beta-cryptogein, which was a difficult target for ab initio predictions in the CASP2 study; it shows that the prediction made with the present method (62% of residues correct) is close to the expected accuracy (66%) for this protein. The larger database and use of a new network training protocol also improve structural class prediction accuracy to 86%, relative to 80% obtained previously. Secondary structure content is predicted with accuracy comparable to that obtained with spectroscopic methods, such as vibrational or electronic circular dichroism and Fourier transform infrared spectroscopy.     

Structure-dependent sequence alignment for remotely related proteins

MOTIVATION: The quality of a model structure derived from a comparative modeling procedure is dictated by the accuracy of the predicted sequence-template alignment. As the sequence-template pairs are increasingly remote in sequence relationship, the prediction of the sequence-template alignments becomes increasingly problematic with sequence alignment methods. Structural information of the template, used in connection with the sequence relationship of the sequence-template pair, could significantly improve the accuracy of the sequence-template alignment. In this paper, we describe a sequence-template alignment method that integrates sequence and structural information to enhance the accuracy of sequence-template alignments for distantly related protein pairs. RESULTS: The structure-dependent sequence alignment (SDSA) procedure was optimized for coverage and accuracy on a training set of 412 protein pairs; the structures for each of the training pairs are similar (RMSD< approximately 4A) but the sequence relationship is undetectable (average pair-wise sequence identity = 8%). The optimized SDSA procedure was then applied to extend PSI-BLAST local alignments by calculating the global alignments under the constraint of the residue pairs in the local alignments. This composite alignment procedure was assessed with a testing set of 1421 protein pairs, of which the pair-wise structures are similar (RMSD< approximately 4A) but the sequences are marginally related at best in each pair (average pair-wise sequence identity = 13%). The assessment showed that the composite alignment procedure predicted more aligned residues pairs with an average of 27% increase in correctly aligned residues over the standard PSI-BLAST alignments for the protein pairs in the testing set.     

Toward an accurate prediction of inter-residue distances in proteins using 2D recursive neural networks

Background

Protein inter-residue contact maps provide a translation and rotation invariant topological representation of a protein. They can be used as an intermediary step in protein structure predictions. However, the prediction of contact maps represents an unbalanced problem as far fewer examples of contacts than non-contacts exist in a protein structure. In this study we explore the possibility of completely eliminating the unbalanced nature of the contact map prediction problem by predicting real-value distances between residues. Predicting full inter-residue distance maps and applying them in protein structure predictions has been relatively unexplored in the past.

Results

We initially demonstrate that the use of native-like distance maps is able to reproduce 3D structures almost identical to the targets, giving an average RMSD of 0.5Å. In addition, the corrupted physical maps with an introduced random error of ±6Å are able to reconstruct the targets within an average RMSD of 2Å. After demonstrating the reconstruction potential of distance maps, we develop two classes of predictors using two-dimensional recursive neural networks: an ab initio predictor that relies only on the protein sequence and evolutionary information, and a template-based predictor in which additional structural homology information is provided. We find that the ab initio predictor is able to reproduce distances with an RMSD of 6Å, regardless of the evolutionary content provided. Furthermore, we show that the template-based predictor exploits both sequence and structure information even in cases of dubious homology and outperforms the best template hit with a clear margin of up to 3.7Å. Lastly, we demonstrate the ability of the two predictors to reconstruct the CASP9 targets shorter than 200 residues producing the results similar to the state of the machine learning art approach implemented in the Distill server.

Conclusions

The methodology presented here, if complemented by more complex reconstruction protocols, can represent a possible path to improve machine learning algorithms for 3D protein structure prediction. Moreover, it can be used as an intermediary step in protein structure predictions either on its own or complemented by NMR restraints.     

Driving Forces for Transmembrane α-Helix Oligomerization

We present what we believe to be a novel statistical contact potential based on solved structures of transmembrane (TM) α-helical bundles, and we use this contact potential to investigate the amino acid likelihood of stabilizing helix-helix interfaces. To increase statistical significance, we have reduced the full contact energy matrix to a four-flavor alphabet of amino acids, automatically determined by our methodology, in which we find that polarity is a more dominant factor of group identity than is size, with charged or polar groups most often occupying the same face, whereas polar/apolar residue pairs tend to occupy opposite faces. We found that the most polar residues strongly influence interhelical contact formation, although they occur rarely in TM helical bundles. Two-body contact energies in the reduced letter code are capable of determining native structure from a large decoy set for a majority of test TM proteins, at the same time illustrating that certain higher-order sequence correlations are necessary for more accurate structure predictions.     

A predictor of transmembrane α-helix domains of proteins based on neural networks

Back-propagation, feed-forward neural networks are used to predict a-helical transmembrane segments of proteins. The networks are trained on the few membrane proteins whose transmembrane -helix domains are known to atomic or nearly atomic resolution. When testing is performed with a jackknife procedure on the proteins of the training set, the fraction of total correct assignments is as high as 0.87, with an average length for the transmembrane segments of 20 residues. The method correctly fails to predict any transmembrane domain for porin, whose transmembrane segments are -sheets. When tested on globular proteins, lower and upper limits of 1.6 and 3.5% for a total of 26826 residues are determined for the mispredicted cases, indicating that the predictor is highly specific for -helical domains of membrane proteins. The predictor is also tested on 37 membrane proteins whose transmembrane topology is partially known. The overall accuracy is 0.90, two percentage points higher than that obtained with statistical methods. The reliability of the prediction is 100% for 60% of the total 18242 predicted residues of membrane proteins. Our results show that the local directional information automatically extracted by the neural networks during the training phase plays a key role in determining the accuracy of the prediction. Correspondence to: R. Casadio     

CNNcon: Improved Protein Contact Maps Prediction Using Cascaded Neural Networks

Backgrounds

Despite continuing progress in X-ray crystallography and high-field NMR spectroscopy for determination of three-dimensional protein structures, the number of unsolved and newly discovered sequences grows much faster than that of determined structures. Protein modeling methods can possibly bridge this huge sequence-structure gap with the development of computational science. A grand challenging problem is to predict three-dimensional protein structure from its primary structure (residues sequence) alone. However, predicting residue contact maps is a crucial and promising intermediate step towards final three-dimensional structure prediction. Better predictions of local and non-local contacts between residues can transform protein sequence alignment to structure alignment, which can finally improve template based three-dimensional protein structure predictors greatly.

Methods

CNNcon, an improved multiple neural networks based contact map predictor using six sub-networks and one final cascade-network, was developed in this paper. Both the sub-networks and the final cascade-network were trained and tested with their corresponding data sets. While for testing, the target protein was first coded and then input to its corresponding sub-networks for prediction. After that, the intermediate results were input to the cascade-network to finish the final prediction.

Results

The CNNcon can accurately predict 58.86% in average of contacts at a distance cutoff of 8 Å for proteins with lengths ranging from 51 to 450. The comparison results show that the present method performs better than the compared state-of-the-art predictors. Particularly, the prediction accuracy keeps steady with the increase of protein sequence length. It indicates that the CNNcon overcomes the thin density problem, with which other current predictors have trouble. This advantage makes the method valuable to the prediction of long length proteins. As a result, the effective prediction of long length proteins could be possible by the CNNcon.     

Residue-level prediction of DNA-binding sites and its application on DNA-binding protein predictions

Protein-DNA interactions are crucial to many cellular activities such as expression-control and DNA-repair. These interactions between amino acids and nucleotides are highly specific and any aberrance at the binding site can render the interaction completely incompetent. In this study, we have three aims focusing on DNA-binding residues on the protein surface: to develop an automated approach for fast and reliable recognition of DNA-binding sites; to improve the prediction by distance-dependent refinement; use these predictions to identify DNA-binding proteins. We use a support vector machines (SVM)-based approach to harness the features of the DNA-binding residues to distinguish them from non-binding residues. Features used for distinction include the residue's identity, charge, solvent accessibility, average potential, the secondary structure it is embedded in, neighboring residues, and location in a cationic patch. These features collected from 50 proteins are used to train SVM. Testing is then performed on another set of 37 proteins, much larger than any testing set used in previous studies. The testing set has no more than 20% sequence identity not only among its pairs, but also with the proteins in the training set, thus removing any undesired redundancy due to homology. This set also has proteins with an unseen DNA-binding structural class not present in the training set. With the above features, an accuracy of 66% with balanced sensitivity and specificity is achieved without relying on homology or evolutionary information. We then develop a post-processing scheme to improve the prediction using the relative location of the predicted residues. Balanced success is then achieved with average sensitivity, specificity and accuracy pegged at 71.3%, 69.3% and 70.5%, respectively. Average net prediction is also around 70%. Finally, we show that the number of predicted DNA-binding residues can be used to differentiate DNA-binding proteins from non-DNA-binding proteins with an accuracy of 78%. Results presented here demonstrate that machine-learning can be applied to automated identification of DNA-binding residues and that the success rate can be ameliorated as more features are added. Such functional site prediction protocols can be useful in guiding consequent works such as site-directed mutagenesis and macromolecular docking.