Improving the prediction of protein secondary structure in three and eight classes using recurrent neural networks and profiles

Secondary structure predictions are increasingly becoming the workhorse for several methods aiming at predicting protein structure and function. Here we use ensembles of bidirectional recurrent neural network architectures, PSI-BLAST-derived profiles, and a large nonredundant training set to derive two new predictors: (a) the second version of the SSpro program for secondary structure classification into three categories and (b) the first version of the SSpro8 program for secondary structure classification into the eight classes produced by the DSSP program. We describe the results of three different test sets on which SSpro achieved a sustained performance of about 78% correct prediction. We report confusion matrices, compare PSI-BLAST to BLAST-derived profiles, and assess the corresponding performance improvements. SSpro and SSpro8 are implemented as web servers, available together with other structural feature predictors at: http://promoter.ics.uci.edu/BRNN-PRED/.     

Porter: a new, accurate server for protein secondary structure prediction

SUMMARY: Porter is a new system for protein secondary structure prediction in three classes. Porter relies on bidirectional recurrent neural networks with shortcut connections, accurate coding of input profiles obtained from multiple sequence alignments, second stage filtering by recurrent neural networks, incorporation of long range information and large-scale ensembles of predictors. Porter's accuracy, tested by rigorous 5-fold cross-validation on a large set of proteins, exceeds 79%, significantly above a copy of the state-of-the-art SSpro server, better than any system published to date. AVAILABILITY: Porter is available as a public web server at http://distill.ucd.ie/porter/ CONTACT: gianluca.pollastri@ucd.ie.     

Distill: a suite of web servers for the prediction of one-, two- and three-dimensional structural features of proteins

Background  

We describe Distill, a suite of servers for the prediction of protein structural features: secondary structure; relative solvent accessibility; contact density; backbone structural motifs; residue contact maps at 6, 8 and 12 Angstrom; coarse protein topology. The servers are based on large-scale ensembles of recursive neural networks and trained on large, up-to-date, non-redundant subsets of the Protein Data Bank. Together with structural feature predictions, Distill includes a server for prediction of C _α traces for short proteins (up to 200 amino acids).     

Strand-loop-strand motifs: prediction of hairpins and diverging turns in proteins

Beta-sheet proteins have been particularly challenging for de novo structure prediction methods, which tend to pair adjacent beta-strands into beta-hairpins and produce overly local topologies. To remedy this problem and facilitate de novo prediction of beta-sheet protein structures, we have developed a neural network that classifies strand-loop-strand motifs by local hairpins and nonlocal diverging turns by using the amino acid sequence as input. The neural network is trained with a representative subset of the Protein Data Bank and achieves a prediction accuracy of 75.9 +/- 4.4% compared to a baseline prediction rate of 59.1%. Hairpins are predicted with an accuracy of 77.3 +/- 6.1%, diverging turns with an accuracy of 73.9 +/- 6.0%. Incorporation of the beta-hairpin/diverging turn classification into the ROSETTA de novo structure prediction method led to higher contact order models and somewhat improved tertiary structure predictions for a test set of 11 all-beta-proteins and 3 alphabeta-proteins. The beta-hairpin/diverging turn classification from amino acid sequences is available online for academic use (Meiler and Kuhn, 2003; www.jens-meiler.de/turnpred.html).     

NetSurfP-2.0: Improved prediction of protein structural features by integrated deep learning

The ability to predict local structural features of a protein from the primary sequence is of paramount importance for unraveling its function in absence of experimental structural information. Two main factors affect the utility of potential prediction tools: their accuracy must enable extraction of reliable structural information on the proteins of interest, and their runtime must be low to keep pace with sequencing data being generated at a constantly increasing speed. Here, we present NetSurfP-2.0, a novel tool that can predict the most important local structural features with unprecedented accuracy and runtime. NetSurfP-2.0 is sequence-based and uses an architecture composed of convolutional and long short-term memory neural networks trained on solved protein structures. Using a single integrated model, NetSurfP-2.0 predicts solvent accessibility, secondary structure, structural disorder, and backbone dihedral angles for each residue of the input sequences. We assessed the accuracy of NetSurfP-2.0 on several independent test datasets and found it to consistently produce state-of-the-art predictions for each of its output features. We observe a correlation of 80% between predictions and experimental data for solvent accessibility, and a precision of 85% on secondary structure 3-class predictions. In addition to improved accuracy, the processing time has been optimized to allow predicting more than 1000 proteins in less than 2 hours, and complete proteomes in less than 1 day.     

Ab initio and template-based prediction of multi-class distance maps by two-dimensional recursive neural networks

Background

Prediction of protein structures from their sequences is still one of the open grand challenges of computational biology. Some approaches to protein structure prediction, especially ab initio ones, rely to some extent on the prediction of residue contact maps. Residue contact map predictions have been assessed at the CASP competition for several years now. Although it has been shown that exact contact maps generally yield correct three-dimensional structures, this is true only at a relatively low resolution (3–4 Å from the native structure). Another known weakness of contact maps is that they are generally predicted ab initio, that is not exploiting information about potential homologues of known structure.

Results

We introduce a new class of distance restraints for protein structures: multi-class distance maps. We show that C _α trace reconstructions based on 4-class native maps are significantly better than those from residue contact maps. We then build two predictors of 4-class maps based on recursive neural networks: one ab initio, or relying on the sequence and on evolutionary information; one template-based, or in which homology information to known structures is provided as a further input. We show that virtually any level of sequence similarity to structural templates (down to less than 10%) yields more accurate 4-class maps than the ab initio predictor. We show that template-based predictions by recursive neural networks are consistently better than the best template and than a number of combinations of the best available templates. We also extract binary residue contact maps at an 8 Å threshold (as per CASP assessment) from the 4-class predictors and show that the template-based version is also more accurate than the best template and consistently better than the ab initio one, down to very low levels of sequence identity to structural templates. Furthermore, we test both ab-initio and template-based 8 Å predictions on the CASP7 targets using a pre-CASP7 PDB, and find that both predictors are state-of-the-art, with the template-based one far outperforming the best CASP7 systems if templates with sequence identity to the query of 10% or better are available. Although this is not the main focus of this paper we also report on reconstructions of C _α traces based on both ab initio and template-based 4-class map predictions, showing that the latter are generally more accurate even when homology is dubious.

Conclusion

Accurate predictions of multi-class maps may provide valuable constraints for improved ab initio and template-based prediction of protein structures, naturally incorporate multiple templates, and yield state-of-the-art binary maps. Predictions of protein structures and 8 Å contact maps based on the multi-class distance map predictors described in this paper are freely available to academic users at the url http://distill.ucd.ie/.     

Modeling protein loops with knowledge-based prediction of sequence-structure alignment

MOTIVATION: As protein structure database expands, protein loop modeling remains an important and yet challenging problem. Knowledge-based protein loop prediction methods have met with two challenges in methodology development: (1) loop boundaries in protein structures are frequently problematic in constructing length-dependent loop databases for protein loop predictions; (2) knowledge-based modeling of loops of unknown structure requires both aligning a query loop sequence to loop templates and ranking the loop sequence-template matches. RESULTS: We developed a knowledge-based loop prediction method that circumvents the need of constructing hierarchically clustered length-dependent loop libraries. The method first predicts local structural fragments of a query loop sequence and then structurally aligns the predicted structural fragments to a set of non-redundant loop structural templates regardless of the loop length. The sequence-template alignments are then quantitatively evaluated with an artificial neural network model trained on a set of predictions with known outcomes. Prediction accuracy benchmarks indicated that the novel procedure provided an alternative approach overcoming the challenges of knowledge-based loop prediction. AVAILABILITY: http://cmb.genomics.sinica.edu.tw     

Beyond the Twilight Zone: Automated prediction of structural properties of proteins by recursive neural networks and remote homology information

The prediction of 1D structural properties of proteins is an important step toward the prediction of protein structure and function, not only in the ab initio case but also when homology information to known structures is available. Despite this the vast majority of 1D predictors do not incorporate homology information into the prediction process. We develop a novel structural alignment method, SAMD, which we use to build alignments of putative remote homologues that we compress into templates of structural frequency profiles. We use these templates as additional input to ensembles of recursive neural networks, which we specialise for the prediction of query sequences that show only remote homology to any Protein Data Bank structure. We predict four 1D structural properties – secondary structure, relative solvent accessibility, backbone structural motifs, and contact density. Secondary structure prediction accuracy, tested by five‐fold cross‐validation on a large set of proteins allowing less than 25% sequence identity between training and test set and query sequences and templates, exceeds 82%, outperforming its ab initio counterpart, other state‐of‐the‐art secondary structure predictors (Jpred 3 and PSIPRED) and two other systems based on PSI‐BLAST and COMPASS templates. We show that structural information from homologues improves prediction accuracy well beyond the Twilight Zone of sequence similarity, even below 5% sequence identity, for all four structural properties. Significant improvement over the extraction of structural information directly from PDB templates suggests that the combination of sequence and template information is more informative than templates alone. Proteins 2009. © 2009 Wiley‐Liss, Inc.     

PaleAle 5.0: prediction of protein relative solvent accessibility by deep learning

Predicting the three-dimensional structure of proteins is a long-standing challenge of computational biology, as the structure (or lack of a rigid structure) is well known to determine a protein’s function. Predicting relative solvent accessibility (RSA) of amino acids within a protein is a significant step towards resolving the protein structure prediction challenge especially in cases in which structural information about a protein is not available by homology transfer. Today, arguably the core of the most powerful prediction methods for predicting RSA and other structural features of proteins is some form of deep learning, and all the state-of-the-art protein structure prediction tools rely on some machine learning algorithm. In this article we present a deep neural network architecture composed of stacks of bidirectional recurrent neural networks and convolutional layers which is capable of mining information from long-range interactions within a protein sequence and apply it to the prediction of protein RSA using a novel encoding method that we shall call “clipped”. The final system we present, PaleAle 5.0, which is available as a public server, predicts RSA into two, three and four classes at an accuracy exceeding 80% in two classes, surpassing the performances of all the other predictors we have benchmarked.

     

Local combinational variables: an approach used in DNA-binding helix-turn-helix motif prediction with sequence information

Sequence-based approach for motif prediction is of great interest and remains a challenge. In this work, we develop a local combinational variable approach for sequence-based helix-turn-helix (HTH) motif prediction. First we choose a sequence data set for 88 proteins of 22 amino acids in length to launch an optimized traversal for extracting local combinational segments (LCS) from the data set. Then after LCS refinement, local combinational variables (LCV) are generated to construct prediction models for HTH motifs. Prediction ability of LCV sets at different thresholds is calculated to settle a moderate threshold. The large data set we used comprises 13 HTH families, with 17 455 sequences in total. Our approach predicts HTH motifs more precisely using only primary protein sequence information, with 93.29% accuracy, 93.93% sensitivity and 92.66% specificity. Prediction results of newly reported HTH-containing proteins compared with other prediction web service presents a good prediction model derived from the LCV approach. Comparisons with profile-HMM models from the Pfam protein families database show that the LCV approach maintains a good balance while dealing with HTH-containing proteins and non-HTH proteins at the same time. The LCV approach is to some extent a complementary to the profile-HMM models for its better identification of false-positive data. Furthermore, genome-wide predictions detect new HTH proteins in both Homo sapiens and Escherichia coli organisms, which enlarge applications of the LCV approach. Software for mining LCVs from sequence data set can be obtained from anonymous ftp site ftp://cheminfo.tongji.edu.cn/LCV/freely.     

Analysis of structural motifs of proteins using sets of codes, describing local structures]

An amino acid sequence pattern conserved among a family of proteins is called motif. It is usually related to the specific function of the family. On the other hand, functions of proteins are achieved by their 3D structures. Specific local structures, called structural motifs, are considered related to their functions. However, searching for common structural motifs in different proteins is much more difficult than for common sequence motifs. We are attempting in this study to convert the information about the structural motifs into a set of one-dimensional digital strings, i.e., a set of codes, to compare them more easily by computer and to investigate their relationship to functions more quantitatively. By applying the Delaunay tessellation to a 3D structure of a protein, we can assign each local structure to a unique code that is defined so as to reflect its structural feature. Since a structural motif is defined as a set of the local structures in this paper, the structural motif is represented by a set of the codes. In order to examine the ability of the set of the codes to distinguish differences among the sets of local structures with a given PROSITE pattern that contain both true and false positives, we clustered them by introducing a similarity measure among the set of the codes. The obtained clustering shows a good agreement with other results by direct structural comparison methods such as a superposition method. The structural motifs in homologous proteins are also properly clustered according to their sources. These results suggest that the structural motifs can be well characterized by these sets of the codes, and that the method can be utilized in comparing structural motifs and relating them with function.     

Prediction of short linear protein binding regions

Short linear motifs in proteins (typically 3-12 residues in length) play key roles in protein-protein interactions by frequently binding specifically to peptide binding domains within interacting proteins. Their tendency to be found in disordered segments of proteins has meant that they have often been overlooked. Here we present SLiMPred (short linear motif predictor), the first general de novo method designed to computationally predict such regions in protein primary sequences independent of experimentally defined homologs and interactors. The method applies machine learning techniques to predict new motifs based on annotated instances from the Eukaryotic Linear Motif database, as well as structural, biophysical, and biochemical features derived from the protein primary sequence. We have integrated these data sources and benchmarked the predictive accuracy of the method, and found that it performs equivalently to a predictor of protein binding regions in disordered regions, in addition to having predictive power for other classes of motif sites such as polyproline II helix motifs and short linear motifs lying in ordered regions. It will be useful in predicting peptides involved in potential protein associations and will aid in the functional characterization of proteins, especially of proteins lacking experimental information on structures and interactions. We conclude that, despite the diversity of motif sequences and structures, SLiMPred is a valuable tool for prioritizing potential interaction motifs in proteins.     

Neuro-fuzzy structural classification of proteins for improved protein secondary structure prediction

Fourier transform infrared (FTIR) spectroscopy is a very flexible technique for characterization of protein secondary structure. Measurements can be carried out rapidly in a number of different environments based on only small quantities of proteins. For this technique to become more widely used for protein secondary structure characterization, however, further developments in methods to accurately quantify protein secondary structure are necessary. Here we propose a structural classification of proteins (SCOP) class specialized neural networks architecture combining an adaptive neuro-fuzzy inference system (ANFIS) with SCOP class specialized backpropagation neural networks for improved protein secondary structure prediction. Our study shows that proteins can be accurately classified into two main classes "all alpha proteins" and "all beta proteins" merely based on the amide I band maximum position of their FTIR spectra. ANFIS is employed to perform the classification task to demonstrate the potential of this architecture with moderately complex problems. Based on studies using a reference set of 17 proteins and an evaluation set of 4 proteins, improved predictions were achieved compared to a conventional neural network approach, where structure specialized neural networks are trained based on protein spectra of both "all alpha" and "all beta" proteins. The standard errors of prediction (SEPs) in % structure were improved by 4.05% for helix structure, by 5.91% for sheet structure, by 2.68% for turn structure, and by 2.15% for bend structure. For other structure, an increase of SEP by 2.43% was observed. Those results were confirmed by a "leave-one-out" run with the combined set of 21 FTIR spectra of proteins.     

Secondary structure prediction with support vector machines

MOTIVATION: A new method that uses support vector machines (SVMs) to predict protein secondary structure is described and evaluated. The study is designed to develop a reliable prediction method using an alternative technique and to investigate the applicability of SVMs to this type of bioinformatics problem. METHODS: Binary SVMs are trained to discriminate between two structural classes. The binary classifiers are combined in several ways to predict multi-class secondary structure. RESULTS: The average three-state prediction accuracy per protein (Q(3)) is estimated by cross-validation to be 77.07 +/- 0.26% with a segment overlap (Sov) score of 73.32 +/- 0.39%. The SVM performs similarly to the 'state-of-the-art' PSIPRED prediction method on a non-homologous test set of 121 proteins despite being trained on substantially fewer examples. A simple consensus of the SVM, PSIPRED and PROFsec achieves significantly higher prediction accuracy than the individual methods.     

Protein contact prediction using patterns of correlation

We describe a new method for using neural networks to predict residue contact pairs in a protein. The main inputs to the neural network are a set of 25 measures of correlated mutation between all pairs of residues in two "windows" of size 5 centered on the residues of interest. While the individual pair-wise correlations are a relatively weak predictor of contact, by training the network on windows of correlation the accuracy of prediction is significantly improved. The neural network is trained on a set of 100 proteins and then tested on a disjoint set of 1033 proteins of known structure. An average predictive accuracy of 21.7% is obtained taking the best L/2 predictions for each protein, where L is the sequence length. Taking the best L/10 predictions gives an average accuracy of 30.7%. The predictor is also tested on a set of 59 proteins from the CASP5 experiment. The accuracy is found to be relatively consistent across different sequence lengths, but to vary widely according to the secondary structure. Predictive accuracy is also found to improve by using multiple sequence alignments containing many sequences to calculate the correlations.     

An empirical energy function for threading protein sequence through the folding motif

In this paper we present a new residue contact potantial derived by statistical analysis of protein crystal structures. This gives mean hydrophobic and pairwise contact energies as a function of residue type and distance interval. To test the accuracy of this potential we generate model structures by “threading” different sequences through backbone folding motifs found in the structural data base. We find that conformational energies calculated by summing contact potentials show perfect specificity in matching the correct sequences with each globular folding motif in a 161-protcin data set. They also identify correct models with the core folding motifs of heme-rythrin and immunoglobulin McPC603 V₁-do- main, among millions of alternatives possible when we align subsequences with α-helices and β-strands, and allow for variation in the lengths of intervening loops. We suggest that contact potentials reflect important constraints on nonbonded interaction in native proteins, and that “threading” may be useful for structure prediction by recognition of folding motif. © 1993 Wiley-Liss, Inc.     

Prediction of the conformation and geometry of loops in globular proteins: testing ArchDB, a structural classification of loops

In protein structure prediction, a central problem is defining the structure of a loop connecting 2 secondary structures. This problem frequently occurs in homology modeling, fold recognition, and in several strategies in ab initio structure prediction. In our previous work, we developed a classification database of structural motifs, ArchDB. The database contains 12,665 clustered loops in 451 structural classes with information about phi-psi angles in the loops and 1492 structural subclasses with the relative locations of the bracing secondary structures. Here we evaluate the extent to which sequence information in the loop database can be used to predict loop structure. Two sequence profiles were used, a HMM profile and a PSSM derived from PSI-BLAST. A jack-knife test was made removing homologous loops using SCOP superfamily definition and predicting afterwards against recalculated profiles that only take into account the sequence information. Two scenarios were considered: (1) prediction of structural class with application in comparative modeling and (2) prediction of structural subclass with application in fold recognition and ab initio. For the first scenario, structural class prediction was made directly over loops with X-ray secondary structure assignment, and if we consider the top 20 classes out of 451 possible classes, the best accuracy of prediction is 78.5%. In the second scenario, structural subclass prediction was made over loops using PSI-PRED (Jones, J Mol Biol 1999;292:195-202) secondary structure prediction to define loop boundaries, and if we take into account the top 20 subclasses out of 1492, the best accuracy is 46.7%. Accuracy of loop prediction was also evaluated by means of RMSD calculations.     

Prediction of protein--protein interaction sites in heterocomplexes with neural networks.

In this paper we address the problem of extracting features relevant for predicting protein--protein interaction sites from the three-dimensional structures of protein complexes. Our approach is based on information about evolutionary conservation and surface disposition. We implement a neural network based system, which uses a cross validation procedure and allows the correct detection of 73% of the residues involved in protein interactions in a selected database comprising 226 heterodimers. Our analysis confirms that the chemico-physical properties of interacting surfaces are difficult to distinguish from those of the whole protein surface. However neural networks trained with a reduced representation of the interacting patch and sequence profile are sufficient to generalize over the different features of the contact patches and to predict whether a residue in the protein surface is or is not in contact. By using a blind test, we report the prediction of the surface interacting sites of three structural components of the Dnak molecular chaperone system, and find close agreement with previously published experimental results. We propose that the predictor can significantly complement results from structural and functional proteomics.