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/.     

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.     

Text Mining for Protein Docking

The rapidly growing amount of publicly available information from biomedical research is readily accessible on the Internet, providing a powerful resource for predictive biomolecular modeling. The accumulated data on experimentally determined structures transformed structure prediction of proteins and protein complexes. Instead of exploring the enormous search space, predictive tools can simply proceed to the solution based on similarity to the existing, previously determined structures. A similar major paradigm shift is emerging due to the rapidly expanding amount of information, other than experimentally determined structures, which still can be used as constraints in biomolecular structure prediction. Automated text mining has been widely used in recreating protein interaction networks, as well as in detecting small ligand binding sites on protein structures. Combining and expanding these two well-developed areas of research, we applied the text mining to structural modeling of protein-protein complexes (protein docking). Protein docking can be significantly improved when constraints on the docking mode are available. We developed a procedure that retrieves published abstracts on a specific protein-protein interaction and extracts information relevant to docking. The procedure was assessed on protein complexes from Dockground (http://dockground.compbio.ku.edu). The results show that correct information on binding residues can be extracted for about half of the complexes. The amount of irrelevant information was reduced by conceptual analysis of a subset of the retrieved abstracts, based on the bag-of-words (features) approach. Support Vector Machine models were trained and validated on the subset. The remaining abstracts were filtered by the best-performing models, which decreased the irrelevant information for ~ 25% complexes in the dataset. The extracted constraints were incorporated in the docking protocol and tested on the Dockground unbound benchmark set, significantly increasing the docking success rate.     

smyRNA: A Novel Ab Initio ncRNA Gene Finder

Background

Non-coding RNAs (ncRNAs) have important functional roles in the cell: for example, they regulate gene expression by means of establishing stable joint structures with target mRNAs via complementary sequence motifs. Sequence motifs are also important determinants of the structure of ncRNAs. Although ncRNAs are abundant, discovering novel ncRNAs on genome sequences has proven to be a hard task; in particular past attempts for ab initio ncRNA search mostly failed with the exception of tools that can identify micro RNAs.

Methodology/Principal Findings

We present a very general ab initio ncRNA gene finder that exploits differential distributions of sequence motifs between ncRNAs and background genome sequences.

Conclusions/Significance

Our method, once trained on a set of ncRNAs from a given species, can be applied to a genome sequences of other organisms to find not only ncRNAs homologous to those in the training set but also others that potentially belong to novel (and perhaps unknown) ncRNA families. Availability: http://compbio.cs.sfu.ca/taverna/smyrna     

Interactome-Wide Prediction of Protein-Protein Binding Sites Reveals Effects of Protein Sequence Variation in Arabidopsis thaliana

The specificity of protein-protein interactions is encoded in those parts of the sequence that compose the binding interface. Therefore, understanding how changes in protein sequence influence interaction specificity, and possibly the phenotype, requires knowing the location of binding sites in those sequences. However, large-scale detection of protein interfaces remains a challenge. Here, we present a sequence- and interactome-based approach to mine interaction motifs from the recently published Arabidopsis thaliana interactome. The resultant proteome-wide predictions are available via www.ab.wur.nl/sliderbio and set the stage for further investigations of protein-protein binding sites. To assess our method, we first show that, by using a priori information calculated from protein sequences, such as evolutionary conservation and residue surface accessibility, we improve the performance of interface prediction compared to using only interactome data. Next, we present evidence for the functional importance of the predicted sites, which are under stronger selective pressure than the rest of protein sequence. We also observe a tendency for compensatory mutations in the binding sites of interacting proteins. Subsequently, we interrogated the interactome data to formulate testable hypotheses for the molecular mechanisms underlying effects of protein sequence mutations. Examples include proteins relevant for various developmental processes. Finally, we observed, by analysing pairs of paralogs, a correlation between functional divergence and sequence divergence in interaction sites. This analysis suggests that large-scale prediction of binding sites can cast light on evolutionary processes that shape protein-protein interaction networks.     

Realistic artificial DNA sequences as negative controls for computational genomics

A common practice in computational genomic analysis is to use a set of ‘background’ sequences as negative controls for evaluating the false-positive rates of prediction tools, such as gene identification programs and algorithms for detection of cis-regulatory elements. Such ‘background’ sequences are generally taken from regions of the genome presumed to be intergenic, or generated synthetically by ‘shuffling’ real sequences. This last method can lead to underestimation of false-positive rates. We developed a new method for generating artificial sequences that are modeled after real intergenic sequences in terms of composition, complexity and interspersed repeat content. These artificial sequences can serve as an inexhaustible source of high-quality negative controls. We used artificial sequences to evaluate the false-positive rates of a set of programs for detecting interspersed repeats, ab initio prediction of coding genes, transcribed regions and non-coding genes. We found that RepeatMasker is more accurate than PClouds, Augustus has the lowest false-positive rate of the coding gene prediction programs tested, and Infernal has a low false-positive rate for non-coding gene detection. A web service, source code and the models for human and many other species are freely available at http://repeatmasker.org/garlic/.     

SHRiMP: Accurate Mapping of Short Color-space Reads

The development of Next Generation Sequencing technologies, capable of sequencing hundreds of millions of short reads (25–70 bp each) in a single run, is opening the door to population genomic studies of non-model species. In this paper we present SHRiMP - the SHort Read Mapping Package: a set of algorithms and methods to map short reads to a genome, even in the presence of a large amount of polymorphism. Our method is based upon a fast read mapping technique, separate thorough alignment methods for regular letter-space as well as AB SOLiD (color-space) reads, and a statistical model for false positive hits. We use SHRiMP to map reads from a newly sequenced Ciona savignyi individual to the reference genome. We demonstrate that SHRiMP can accurately map reads to this highly polymorphic genome, while confirming high heterozygosity of C. savignyi in this second individual. SHRiMP is freely available at http://compbio.cs.toronto.edu/shrimp.     

Combining Evolutionary Information and an Iterative Sampling Strategy for Accurate Protein Structure Prediction

Recent work has shown that the accuracy of ab initio structure prediction can be significantly improved by integrating evolutionary information in form of intra-protein residue-residue contacts. Following this seminal result, much effort is put into the improvement of contact predictions. However, there is also a substantial need to develop structure prediction protocols tailored to the type of restraints gained by contact predictions. Here, we present a structure prediction protocol that combines evolutionary information with the resolution-adapted structural recombination approach of Rosetta, called RASREC. Compared to the classic Rosetta ab initio protocol, RASREC achieves improved sampling, better convergence and higher robustness against incorrect distance restraints, making it the ideal sampling strategy for the stated problem. To demonstrate the accuracy of our protocol, we tested the approach on a diverse set of 28 globular proteins. Our method is able to converge for 26 out of the 28 targets and improves the average TM-score of the entire benchmark set from 0.55 to 0.72 when compared to the top ranked models obtained by the EVFold web server using identical contact predictions. Using a smaller benchmark, we furthermore show that the prediction accuracy of our method is only slightly reduced when the contact prediction accuracy is comparatively low. This observation is of special interest for protein sequences that only have a limited number of homologs.     

Polyploidy and novelty: Gottlieb's legacy

Nearly four decades ago, Roose & Gottlieb (Roose & Gottlieb 1976 Evolution 30, 818–830. (doi:10.2307/2407821)) showed that the recently derived allotetraploids Tragopogon mirus and T. miscellus combined the allozyme profiles of their diploid parents (T. dubius and T. porrifolius, and T. dubius and T. pratensis, respectively). This classic paper addressed the link between genotype and biochemical phenotype and documented enzyme additivity in allopolyploids. Perhaps more important than their model of additivity, however, was their demonstration of novelty at the biochemical level. Enzyme multiplicity—the production of novel enzyme forms in the allopolyploids—can provide an extensive array of polymorphism for a polyploid individual and may explain, for example, the expanded ranges of polyploids relative to their diploid progenitors. In this paper, we extend the concept of evolutionary novelty in allopolyploids to a range of genetic and ecological features. We observe that the dynamic nature of polyploid genomes—with alterations in gene content, gene number, gene arrangement, gene expression and transposon activity—may generate sufficient novelty that every individual in a polyploid population or species may be unique. Whereas certain combinations of these features will undoubtedly be maladaptive, some unique combinations of newly generated variation may provide tremendous evolutionary potential and adaptive capabilities.     

THetA: inferring intra-tumor heterogeneity from high-throughput DNA sequencing data

Tumor samples are typically heterogeneous, containing admixture by normal, non-cancerous cells and one or more subpopulations of cancerous cells. Whole-genome sequencing of a tumor sample yields reads from this mixture, but does not directly reveal the cell of origin for each read. We introduce THetA (Tumor Heterogeneity Analysis), an algorithm that infers the most likely collection of genomes and their proportions in a sample, for the case where copy number aberrations distinguish subpopulations. THetA successfully estimates normal admixture and recovers clonal and subclonal copy number aberrations in real and simulated sequencing data. THetA is available at http://compbio.cs.brown.edu/software/.     

CPORT: a consensus interface predictor and its performance in prediction-driven docking with HADDOCK

Background

Macromolecular complexes are the molecular machines of the cell. Knowledge at the atomic level is essential to understand and influence their function. However, their number is huge and a significant fraction is extremely difficult to study using classical structural methods such as NMR and X-ray crystallography. Therefore, the importance of large-scale computational approaches in structural biology is evident. This study combines two of these computational approaches, interface prediction and docking, to obtain atomic-level structures of protein-protein complexes, starting from their unbound components.

Methodology/Principal Findings

Here we combine six interface prediction web servers into a consensus method called CPORT (Consensus Prediction Of interface Residues in Transient complexes). We show that CPORT gives more stable and reliable predictions than each of the individual predictors on its own. A protocol was developed to integrate CPORT predictions into our data-driven docking program HADDOCK. For cases where experimental information is limited, this prediction-driven docking protocol presents an alternative to ab initio docking, the docking of complexes without the use of any information. Prediction-driven docking was performed on a large and diverse set of protein-protein complexes in a blind manner. Our results indicate that the performance of the HADDOCK-CPORT combination is competitive with ZDOCK-ZRANK, a state-of-the-art ab initio docking/scoring combination. Finally, the original interface predictions could be further improved by interface post-prediction (contact analysis of the docking solutions).

Conclusions/Significance

The current study shows that blind, prediction-driven docking using CPORT and HADDOCK is competitive with ab initio docking methods. This is encouraging since prediction-driven docking represents the absolute bottom line for data-driven docking: any additional biological knowledge will greatly improve the results obtained by prediction-driven docking alone. Finally, the fact that original interface predictions could be further improved by interface post-prediction suggests that prediction-driven docking has not yet been pushed to the limit. A web server for CPORT is freely available at http://haddock.chem.uu.nl/services/CPORT.     

IPred - integrating ab initio and evidence based gene predictions to improve prediction accuracy

Background

Gene prediction is a challenging but crucial part in most genome analysis pipelines. Various methods have evolved that predict genes ab initio on reference sequences or evidence based with the help of additional information, such as RNA-Seq reads or EST libraries. However, none of these strategies is bias-free and one method alone does not necessarily provide a complete set of accurate predictions.

Results

We present IPred (Integrative gene Prediction), a method to integrate ab initio and evidence based gene identifications to complement the advantages of different prediction strategies. IPred builds on the output of gene finders and generates a new combined set of gene identifications, representing the integrated evidence of the single method predictions.

Conclusion

We evaluate IPred in simulations and real data experiments on Escherichia Coli and human data. We show that IPred improves the prediction accuracy in comparison to single method predictions and to existing methods for prediction combination.

Electronic supplementary material

The online version of this article (doi:10.1186/s12864-015-1315-9) contains supplementary material, which is available to authorized users.     

PSPP: A Protein Structure Prediction Pipeline for Computing Clusters

Background

Protein structures are critical for understanding the mechanisms of biological systems and, subsequently, for drug and vaccine design. Unfortunately, protein sequence data exceed structural data by a factor of more than 200 to 1. This gap can be partially filled by using computational protein structure prediction. While structure prediction Web servers are a notable option, they often restrict the number of sequence queries and/or provide a limited set of prediction methodologies. Therefore, we present a standalone protein structure prediction software package suitable for high-throughput structural genomic applications that performs all three classes of prediction methodologies: comparative modeling, fold recognition, and ab initio. This software can be deployed on a user''s own high-performance computing cluster.

Methodology/Principal Findings

The pipeline consists of a Perl core that integrates more than 20 individual software packages and databases, most of which are freely available from other research laboratories. The query protein sequences are first divided into domains either by domain boundary recognition or Bayesian statistics. The structures of the individual domains are then predicted using template-based modeling or ab initio modeling. The predicted models are scored with a statistical potential and an all-atom force field. The top-scoring ab initio models are annotated by structural comparison against the Structural Classification of Proteins (SCOP) fold database. Furthermore, secondary structure, solvent accessibility, transmembrane helices, and structural disorder are predicted. The results are generated in text, tab-delimited, and hypertext markup language (HTML) formats. So far, the pipeline has been used to study viral and bacterial proteomes.

Conclusions

The standalone pipeline that we introduce here, unlike protein structure prediction Web servers, allows users to devote their own computing assets to process a potentially unlimited number of queries as well as perform resource-intensive ab initio structure prediction.