The pros and cons of predicting protein contact maps

MBE: model-based enrichment estimation and prediction for differential sequencing data

Characterizing differences in sequences between two conditions, such as with and without drug exposure, using high-throughput sequencing data is a prevalent problem involving quantifying changes in sequence abundances, and predicting such differences for unobserved sequences. A key shortcoming of current approaches is their extremely limited ability to share information across related but non-identical reads. Consequently, they cannot use sequencing data effectively, nor be directly applied in many settings of interest. We introduce model-based enrichment (MBE) to overcome this shortcoming. We evaluate MBE using both simulated and real data. Overall, MBE improves accuracy compared to current differential analysis methods.

Alignments of biomolecular contact maps

Alignments of discrete objects can be constructed in a very general setting as super-objects from which the constituent objects are recovered by means of projections. Here, we focus on contact maps, i.e. undirected graphs with an ordered set of vertices. These serve as natural discretizations of RNA and protein structures. In the general case, the alignment problem for vertex-ordered graphs is NP-complete. In the special case of RNA secondary structures, i.e. crossing-free matchings, however, the alignments have a recursive structure. The alignment problem then can be solved by a variant of the Sankoff algorithm in polynomial time. Moreover, the tree or forest alignments of RNA secondary structure can be understood as the alignments of ordered edge sets.

Unsupervised protein embeddings outperform hand-crafted sequence and structure features at predicting molecular function

MOTIVATION

Protein function prediction is a difficult bioinformatics problem. Many recent methods use deep neural networks to learn complex sequence representations and predict function from these. Deep supervised models require a lot of labeled training data which are not available for this task. However, a very large amount of protein sequences without functional labels is available.

RESULTS

We applied an existing deep sequence model that had been pretrained in an unsupervised setting on the supervised task of protein molecular function prediction. We found that this complex feature representation is effective for this task, outperforming hand-crafted features such as one-hot encoding of amino acids, k-mer counts, secondary structure and backbone angles. Also, it partly negates the need for complex prediction models, as a two-layer perceptron was enough to achieve competitive performance in the third Critical Assessment of Functional Annotation benchmark. We also show that combining this sequence representation with protein 3D structure information does not lead to performance improvement, hinting that 3D structure is also potentially learned during the unsupervised pretraining.

AVAILABILITY AND IMPLEMENTATION

Implementations of all used models can be found at https://github.com/stamakro/GCN-for-Structure-and-Function.

SUPPLEMENTARY INFORMATION

Supplementary data are available at Bioinformatics online.

Deep learning methods in protein structure prediction

Protein Structure Prediction is a central topic in Structural Bioinformatics. Since the '60s statistical methods, followed by increasingly complex Machine Learning and recently Deep Learning methods, have been employed to predict protein structural information at various levels of detail. In this review, we briefly introduce the problem of protein structure prediction and essential elements of Deep Learning (such as Convolutional Neural Networks, Recurrent Neural Networks and basic feed-forward Neural Networks they are founded on), after which we discuss the evolution of predictive methods for one-dimensional and two-dimensional Protein Structure Annotations, from the simple statistical methods of the early days, to the computationally intensive highly-sophisticated Deep Learning algorithms of the last decade. In the process, we review the growth of the databases these algorithms are based on, and how this has impacted our ability to leverage knowledge about evolution and co-evolution to achieve improved predictions. We conclude this review outlining the current role of Deep Learning techniques within the wider pipelines to predict protein structures and trying to anticipate what challenges and opportunities may arise next.

StructureDistiller: Structural relevance scoring identifies the most informative entries of a contact map

Protein folding and structure prediction are two sides of the same coin. Contact maps and the related techniques of constraint-based structure reconstruction can be considered as unifying aspects of both processes. We present the Structural Relevance (SR) score which quantifies the information content of individual contacts and residues in the context of the whole native structure. The physical process of protein folding is commonly characterized with spatial and temporal resolution: some residues are Early Folding while others are Highly Stable with respect to unfolding events. We employ the proposed SR score to demonstrate that folding initiation and structure stabilization are subprocesses realized by distinct sets of residues. The example of cytochrome c is used to demonstrate how StructureDistiller identifies the most important contacts needed for correct protein folding. This shows that entries of a contact map are not equally relevant for structural integrity. The proposed StructureDistiller algorithm identifies contacts with the highest information content; these entries convey unique constraints not captured by other contacts. Identification of the most informative contacts effectively doubles resilience toward contacts which are not observed in the native contact map. Furthermore, this knowledge increases reconstruction fidelity on sparse contact maps significantly by 0.4 Å.

ConSole: using modularity of contact maps to locate solenoid domains in protein structures

Background

Periodic proteins, characterized by the presence of multiple repeats of short motifs, form an interesting and seldom-studied group. Due to often extreme divergence in sequence, detection and analysis of such motifs is performed more reliably on the structural level. Yet, few algorithms have been developed for the detection and analysis of structures of periodic proteins.

Results

ConSole recognizes modularity in protein contact maps, allowing for precise identification of repeats in solenoid protein structures, an important subgroup of periodic proteins. Tests on benchmarks show that ConSole has higher recognition accuracy as compared to Raphael, the only other publicly available solenoid structure detection tool. As a next step of ConSole analysis, we show how detection of solenoid repeats in structures can be used to improve sequence recognition of these motifs and to detect subtle irregularities of repeat lengths in three solenoid protein families.

Conclusions

The ConSole algorithm provides a fast and accurate tool to recognize solenoid protein structures as a whole and to identify individual solenoid repeat units from a structure. ConSole is available as a web-based, interactive server and is available for download at http://console.sanfordburnham.org.

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.

Community structure description in amino acid interaction networks

In this paper, we represent proteins by amino acid interaction networks. This is a graph whose vertices are the protein's amino acids and whose edges are the interactions between them. We begin by identifying the main topological properties of these interaction networks using graph theory measures. We observe that the amino acids interact specifically, according to their structural role, and depending on whether they participate or not in the secondary structure. Thus, certain amino acids tend to group together to form local clouds. Then, we study the formation of node aggregations through community structure detections. We observe that the composition of organizations confirms a specific aggregation between loops around a core composed of secondary.

Blurring contact maps of thousands of proteins: what we can learn by reconstructing 3D structure

Background

The present knowledge of protein structures at atomic level derives from some 60,000 molecules. Yet the exponential ever growing set of hypothetical protein sequences comprises some 10 million chains and this makes the problem of protein structure prediction one of the challenging goals of bioinformatics. In this context, the protein representation with contact maps is an intermediate step of fold recognition and constitutes the input of contact map predictors. However contact map representations require fast and reliable methods to reconstruct the specific folding of the protein backbone.

Methods

In this paper, by adopting a GRID technology, our algorithm for 3D reconstruction FT-COMAR is benchmarked on a huge set of non redundant proteins (1716) taking random noise into consideration and this makes our computation the largest ever performed for the task at hand.

Results

We can observe the effects of introducing random noise on 3D reconstruction and derive some considerations useful for future implementations. The dimension of the protein set allows also statistical considerations after grouping per SCOP structural classes.

Conclusions

All together our data indicate that the quality of 3D reconstruction is unaffected by deleting up to an average 75% of the real contacts while only few percentage of randomly generated contacts in place of non-contacts are sufficient to hamper 3D reconstruction.

Optimal contact definition for reconstruction of contact maps

Background

Contact maps have been extensively used as a simplified representation of protein structures. They capture most important features of a protein's fold, being preferred by a number of researchers for the description and study of protein structures. Inspired by the model's simplicity many groups have dedicated a considerable amount of effort towards contact prediction as a proxy for protein structure prediction. However a contact map's biological interest is subject to the availability of reliable methods for the 3-dimensional reconstruction of the structure.

Results

We use an implementation of the well-known distance geometry protocol to build realistic protein 3-dimensional models from contact maps, performing an extensive exploration of many of the parameters involved in the reconstruction process. We try to address the questions: a) to what accuracy does a contact map represent its corresponding 3D structure, b) what is the best contact map representation with regard to reconstructability and c) what is the effect of partial or inaccurate contact information on the 3D structure recovery. Our results suggest that contact maps derived from the application of a distance cutoff of 9 to 11Å around the C_β atoms constitute the most accurate representation of the 3D structure. The reconstruction process does not provide a single solution to the problem but rather an ensemble of conformations that are within 2Å RMSD of the crystal structure and with lower values for the pairwise average ensemble RMSD. Interestingly it is still possible to recover a structure with partial contact information, although wrong contacts can lead to dramatic loss in reconstruction fidelity.

Conclusions

Thus contact maps represent a valid approximation to the structures with an accuracy comparable to that of experimental methods. The optimal contact definitions constitute key guidelines for methods based on contact maps such as structure prediction through contacts and structural alignments based on maximum contact map overlap.

Atomic interaction networks in the core of protein domains and their native folds

Vastly divergent sequences populate a majority of protein folds. In the quest to identify features that are conserved within protein domains belonging to the same fold, we set out to examine the entire protein universe on a fold-by-fold basis. We report that the atomic interaction network in the solvent-unexposed core of protein domains are fold-conserved, extraordinary sequence divergence notwithstanding. Further, we find that this feature, termed protein core atomic interaction network (or PCAIN) is significantly distinguishable across different folds, thus appearing to be “signature” of a domain's native fold. As part of this study, we computed the PCAINs for 8698 representative protein domains from families across the 1018 known protein folds to construct our seed database and an automated framework was developed for PCAIN-based characterization of the protein fold universe. A test set of randomly selected domains that are not in the seed database was classified with over 97% accuracy, independent of sequence divergence. As an application of this novel fold signature, a PCAIN-based scoring scheme was developed for comparative (homology-based) structure prediction, with 1–2 angstroms (mean 1.61A) C_α RMSD generally observed between computed structures and reference crystal structures. Our results are consistent across the full spectrum of test domains including those from recent CASP experiments and most notably in the ‘twilight’ and ‘midnight’ zones wherein <30% and <10% target-template sequence identity prevails (mean twilight RMSD of 1.69A). We further demonstrate the utility of the PCAIN protocol to derive biological insight into protein structure-function relationships, by modeling the structure of the YopM effector novel E3 ligase (NEL) domain from plague-causative bacterium Yersinia Pestis and discussing its implications for host adaptive and innate immune modulation by the pathogen. Considering the several high-throughput, sequence-identity-independent applications demonstrated in this work, we suggest that the PCAIN is a fundamental fold feature that could be a valuable addition to the arsenal of protein modeling and analysis tools.