Prediction of contact maps with neural networks and correlated mutations

The physics-AI dialogue in drug design

A long path has led from the determination of the first protein structure in 1960 to the recent breakthroughs in protein science. Protein structure prediction and design methodologies based on machine learning (ML) have been recognized with the 2024 Nobel prize in Chemistry, but they would not have been possible without previous work and the input of many domain scientists. Challenges remain in the application of ML tools for the prediction of structural ensembles and their usage within the software pipelines for structure determination by crystallography or cryogenic electron microscopy. In the drug discovery workflow, ML techniques are being used in diverse areas such as scoring of docked poses, or the generation of molecular descriptors. As the ML techniques become more widespread, novel applications emerge which can profit from the large amounts of data available. Nevertheless, it is essential to balance the potential advantages against the environmental costs of ML deployment to decide if and when it is best to apply it. For hit to lead optimization ML tools can efficiently interpolate between compounds in large chemical series but free energy calculations by molecular dynamics simulations seem to be superior for designing novel derivatives. Importantly, the potential complementarity and/or synergism of physics-based methods (e.g., force field-based simulation models) and data-hungry ML techniques is growing strongly. Current ML methods have evolved from decades of research. It is now necessary for biologists, physicists, and computer scientists to fully understand advantages and limitations of ML techniques to ensure that the complementarity of physics-based methods and ML tools can be fully exploited for drug design.

Dual modality feature fused neural network integrating binding site information for drug target affinity prediction

Accurately predicting binding affinities between drugs and targets is crucial for drug discovery but remains challenging due to the complexity of modeling interactions between small drug and large targets. This study proposes DMFF-DTA, a dual-modality neural network model integrates sequence and graph structure information from drugs and proteins for drug-target affinity prediction. The model introduces a binding site-focused graph construction approach to extract binding information, enabling more balanced and efficient modeling of drug-target interactions. Comprehensive experiments demonstrate DMFF-DTA outperforms state-of-the-art methods with significant improvements. The model exhibits excellent generalization capabilities on completely unseen drugs and targets, achieving an improvement of over 8% compared to existing methods. Model interpretability analysis validates the biological relevance of the model. A case study in pancreatic cancer drug repurposing demonstrates its practical utility. This work provides an interpretable, robust approach to integrate multi-view drug and protein features for advancing computational drug discovery.

AI in cellular engineering and reprogramming

During the last decade, artificial intelligence (AI) has increasingly been applied in biophysics and related fields, including cellular engineering and reprogramming, offering novel approaches to understand, manipulate, and control cellular function. The potential of AI lies in its ability to analyze complex datasets and generate predictive models. AI algorithms can process large amounts of data from single-cell genomics and multiomic technologies, allowing researchers to gain mechanistic insights into the control of cell identity and function. By integrating and interpreting these complex datasets, AI can help identify key molecular events and regulatory pathways involved in cellular reprogramming. This knowledge can inform the design of precision engineering strategies, such as the development of new transcription factor and signaling molecule cocktails, to manipulate cell identity and drive authentic cell fate across lineage boundaries. Furthermore, when used in combination with computational methods, AI can accelerate and improve the analysis and understanding of the intricate relationships between genes, proteins, and cellular processes. In this review article, we explore the current state of AI applications in biophysics with a specific focus on cellular engineering and reprogramming. Then, we showcase a couple of recent applications where we combined machine learning with experimental and computational techniques. Finally, we briefly discuss the challenges and prospects of AI in cellular engineering and reprogramming, emphasizing the potential of these technologies to revolutionize our ability to engineer cells for a variety of applications, from disease modeling and drug discovery to regenerative medicine and biomanufacturing.

Harnessing the power of artificial intelligence to advance cell therapy

Cell therapies are powerful technologies in which human cells are reprogrammed for therapeutic applications such as killing cancer cells or replacing defective cells. The technologies underlying cell therapies are increasing in effectiveness and complexity, making rational engineering of cell therapies more difficult. Creating the next generation of cell therapies will require improved experimental approaches and predictive models. Artificial intelligence (AI) and machine learning (ML) methods have revolutionized several fields in biology including genome annotation, protein structure prediction, and enzyme design. In this review, we discuss the potential of combining experimental library screens and AI to build predictive models for the development of modular cell therapy technologies. Advances in DNA synthesis and high-throughput screening techniques enable the construction and screening of libraries of modular cell therapy constructs. AI and ML models trained on this screening data can accelerate the development of cell therapies by generating predictive models, design rules, and improved designs.

Protein folding problem: enigma, paradox, solution

The ability of protein chains to spontaneously form their three-dimensional structures is a long-standing mystery in molecular biology. The most conceptual aspect of this mystery is how the protein chain can find its native, "working" spatial structure (which, for not too big protein chains, corresponds to the global free energy minimum) in a biologically reasonable time, without exhaustive enumeration of all possible conformations, which would take billions of years. This is the so-called "Levinthal's paradox." In this review, we discuss the key ideas and discoveries leading to the current understanding of protein folding kinetics, including folding landscapes and funnels, free energy barriers at the folding/unfolding pathways, and the solution of Levinthal's paradox. A special role here is played by the "all-or-none" phase transition occurring at protein folding and unfolding and by the point of thermodynamic (and kinetic) equilibrium between the "native" and the "unfolded" phases of the protein chain (where the theory obtains the simplest form). The modern theory provides an understanding of key features of protein folding and, in good agreement with experiments, it (i) outlines the chain length-dependent range of protein folding times, (ii) predicts the observed maximal size of "foldable" proteins and domains. Besides, it predicts the maximal size of proteins and domains that fold under solely thermodynamic (rather than kinetic) control. Complementarily, a theoretical analysis of the number of possible protein folding patterns, performed at the level of formation and assembly of secondary structures, correctly outlines the upper limit of protein folding times.

Glycoinformatics in the Artificial Intelligence Era

Artificial intelligence (AI) methods have been and are now being increasingly integrated in prediction software implemented in bioinformatics and its glycoscience branch known as glycoinformatics. AI techniques have evolved in the past decades, and their applications in glycoscience are not yet widespread. This limited use is partly explained by the peculiarities of glyco-data that are notoriously hard to produce and analyze. Nonetheless, as time goes, the accumulation of glycomics, glycoproteomics, and glycan-binding data has reached a point where even the most recent deep learning methods can provide predictors with good performance. We discuss the historical development of the application of various AI methods in the broader field of glycoinformatics. A particular focus is placed on shining a light on challenges in glyco-data handling, contextualized by lessons learnt from related disciplines. Ending on the discussion of state-of-the-art deep learning approaches in glycoinformatics, we also envision the future of glycoinformatics, including development that need to occur in order to truly unleash the capabilities of glycoscience in the systems biology era.

Enhancing protein inter-residue real distance prediction by scrutinising deep learning models

Protein structure prediction (PSP) has achieved significant progress lately via prediction of inter-residue distances using deep learning models and exploitation of the predictions during conformational search. In this context, prediction of large inter-residue distances and also prediction of distances between residues separated largely in the protein sequence remain challenging. To deal with these challenges, state-of-the-art inter-residue distance prediction algorithms have used large sets of coevolutionary and non-coevolutionary features. In this paper, we argue that the more the types of features used, the more the kinds of noises introduced and then the deep learning model has to overcome the noises to improve the accuracy of the predictions. Also, multiple features capturing similar underlying characteristics might not necessarily have significantly better cumulative effect. So we scrutinise the feature space to reduce the types of features to be used, but at the same time, we strive to improve the prediction accuracy. Consequently, for inter-residue real distance prediction, in this paper, we propose a deep learning model named scrutinised distance predictor (SDP), which uses only 2 coevolutionary and 3 non-coevolutionary features. On several sets of benchmark proteins, our proposed SDP method improves mean Local Distance Different Test (LDDT) scores at least by 10% over existing state-of-the-art methods. The SDP program along with its data is available from the website https://gitlab.com/mahnewton/sdp .

Boltzmann Machine Learning and Regularization Methods for Inferring Evolutionary Fields and Couplings From a Multiple Sequence Alignment

The inverse Potts problem to infer a Boltzmann distribution for homologous protein sequences from their single-site and pairwise amino acid frequencies recently attracts a great deal of attention in the studies of protein structure and evolution. We study regularization and learning methods and how to tune regularization parameters to correctly infer interactions in Boltzmann machine learning. Using L₂ regularization for fields, group L₁ for couplings is shown to be very effective for sparse couplings in comparison with L₂ and L₁. Two regularization parameters are tuned to yield equal values for both the sample and ensemble averages of evolutionary energy. Both averages smoothly change and converge, but their learning profiles are very different between learning methods. The Adam method is modified to make stepsize proportional to the gradient for sparse couplings and to use a soft-thresholding function for group L₁. It is shown by first inferring interactions from protein sequences and then from Monte Carlo samples that the fields and couplings can be well recovered, but that recovering the pairwise correlations in the resolution of a total energy is harder for the natural proteins than for the protein-like sequences. Selective temperature for folding/structural constrains in protein evolution is also estimated.

Highly accurate protein structure prediction with AlphaFold

Proteins are essential to life, and understanding their structure can facilitate a mechanistic understanding of their function. Through an enormous experimental effort1-4, the structures of around 100,000 unique proteins have been determined5, but this represents a small fraction of the billions of known protein sequences6,7. Structural coverage is bottlenecked by the months to years of painstaking effort required to determine a single protein structure. Accurate computational approaches are needed to address this gap and to enable large-scale structural bioinformatics. Predicting the three-dimensional structure that a protein will adopt based solely on its amino acid sequence-the structure prediction component of the 'protein folding problem'8-has been an important open research problem for more than 50 years9. Despite recent progress10-14, existing methods fall far short of atomic accuracy, especially when no homologous structure is available. Here we provide the first computational method that can regularly predict protein structures with atomic accuracy even in cases in which no similar structure is known. We validated an entirely redesigned version of our neural network-based model, AlphaFold, in the challenging 14th Critical Assessment of protein Structure Prediction (CASP14)15, demonstrating accuracy competitive with experimental structures in a majority of cases and greatly outperforming other methods. Underpinning the latest version of AlphaFold is a novel machine learning approach that incorporates physical and biological knowledge about protein structure, leveraging multi-sequence alignments, into the design of the deep learning algorithm.

Toward the solution of the protein structure prediction problem

Since Anfinsen demonstrated that the information encoded in a protein's amino acid sequence determines its structure in 1973, solving the protein structure prediction problem has been the Holy Grail of structural biology. The goal of protein structure prediction approaches is to utilize computational modeling to determine the spatial location of every atom in a protein molecule starting from only its amino acid sequence. Depending on whether homologous structures can be found in the Protein Data Bank (PDB), structure prediction methods have been historically categorized as template-based modeling (TBM) or template-free modeling (FM) approaches. Until recently, TBM has been the most reliable approach to predicting protein structures, and in the absence of reliable templates, the modeling accuracy sharply declines. Nevertheless, the results of the most recent community-wide assessment of protein structure prediction experiment (CASP14) have demonstrated that the protein structure prediction problem can be largely solved through the use of end-to-end deep machine learning techniques, where correct folds could be built for nearly all single-domain proteins without using the PDB templates. Critically, the model quality exhibited little correlation with the quality of available template structures, as well as the number of sequence homologs detected for a given target protein. Thus, the implementation of deep-learning techniques has essentially broken through the 50-year-old modeling border between TBM and FM approaches and has made the success of high-resolution structure prediction significantly less dependent on template availability in the PDB library.

Protein Contact Map Refinement for Improving Structure Prediction Using Generative Adversarial Networks

MOTIVATION

Protein structure prediction remains as one of the most important problems in computational biology and biophysics. In the past few years, protein residue-residue contact prediction has undergone substantial improvement, which has made it a critical driving force for successful protein structure prediction. Boosting the accuracy of contact predictions has, therefore, become the forefront of protein structure prediction.

RESULTS

We show a novel contact map refinement method, ContactGAN, which uses Generative Adversarial Networks (GAN). ContactGAN was able to make a significant improvement over predictions made by recent contact prediction methods when tested on three datasets including protein structure modeling targets in CASP13 and CASP14. We show improvement of precision in contact prediction, which translated into improvement in the accuracy of protein tertiary structure models. On the other hand, observed improvement over trRosetta was relatively small, reasons for which are discussed. ContactGAN will be a valuable addition in the structure prediction pipeline to achieve an extra gain in contact prediction accuracy.

AVAILABILITY

https://github.com/kiharalab/ContactGAN.

SUPPLEMENTARY INFORMATION

Supplementary data are available at Bioinformatics online.

Deep Learning in Protein Structural Modeling and Design

Deep learning is catalyzing a scientific revolution fueled by big data, accessible toolkits, and powerful computational resources, impacting many fields, including protein structural modeling. Protein structural modeling, such as predicting structure from amino acid sequence and evolutionary information, designing proteins toward desirable functionality, or predicting properties or behavior of a protein, is critical to understand and engineer biological systems at the molecular level. In this review, we summarize the recent advances in applying deep learning techniques to tackle problems in protein structural modeling and design. We dissect the emerging approaches using deep learning techniques for protein structural modeling and discuss advances and challenges that must be addressed. We argue for the central importance of structure, following the "sequence → structure → function" paradigm. This review is directed to help both computational biologists to gain familiarity with the deep learning methods applied in protein modeling, and computer scientists to gain perspective on the biologically meaningful problems that may benefit from deep learning techniques.

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.

ProDCoNN: Protein design using a convolutional neural network

Designing protein sequences that fold to a given three-dimensional (3D) structure has long been a challenging problem in computational structural biology with significant theoretical and practical implications. In this study, we first formulated this problem as predicting the residue type given the 3D structural environment around the C _α atom of a residue, which is repeated for each residue of a protein. We designed a nine-layer 3D deep convolutional neural network (CNN) that takes as input a gridded box with the atomic coordinates and types around a residue. Several CNN layers were designed to capture structure information at different scales, such as bond lengths, bond angles, torsion angles, and secondary structures. Trained on a very large number of protein structures, the method, called ProDCoNN (protein design with CNN), achieved state-of-the-art performance when tested on large numbers of test proteins and benchmark datasets.

Assessing the accuracy of contact predictions in CASP13

The accuracy of sequence-based tertiary contact predictions was assessed in a blind prediction experiment at the CASP13 meeting. After 4 years of significant improvements in prediction accuracy, another dramatic advance has taken place since CASP12 was held 2 years ago. The precision of predicting the top L/5 contacts in the free modeling category, where L is the corresponding length of the protein in residues, has exceeded 70%. As a comparison, the best-performing group at CASP12 with a 47% precision would have finished below the top 1/3 of the CASP13 groups. Extensively trained deep neural network approaches dominate the top performing algorithms, which appear to efficiently integrate information on coevolving residues and interacting fragments or possibly utilize memories of sequence similarities and sometimes can deliver accurate results even in the absence of virtually any target specific evolutionary information. If the current performance is evaluated by F-score on L contacts, it stands around 24% right now, which, despite the tremendous impact and advance in improving its utility for structure modeling, also suggests that there is much room left for further improvement.

Bio-knowledge-based filters improve residue-residue contact prediction accuracy

Motivation

Residue-residue contact prediction through direct coupling analysis has reached impressive accuracy, but yet higher accuracy will be needed to allow for routine modelling of protein structures. One way to improve the prediction accuracy is to filter predicted contacts using knowledge about the particular protein of interest or knowledge about protein structures in general.

Results

We focus on the latter and discuss a set of filters that can be used to remove false positive contact predictions. Each filter depends on one or a few cut-off parameters for which the filter performance was investigated. Combining all filters while using default parameters resulted for a test set of 851 protein domains in the removal of 29% of the predictions of which 92% were indeed false positives.

Availability and implementation

All data and scripts are available at http://comprec-lin.iiar.pwr.edu.pl/FPfilter/.

Supplementary information

Supplementary data are available at Bioinformatics online.

Protein Inter-Residue Contacts Prediction: Methods, Performances and Applications

Background:

Protein inter-residue contacts prediction play an important role in the field of protein structure and function research. As a low-dimensional representation of protein tertiary structure, protein inter-residue contacts could greatly help de novo protein structure prediction methods to reduce the conformational search space. Over the past two decades, various methods have been developed for protein inter-residue contacts prediction.

Objective:

We provide a comprehensive and systematic review of protein inter-residue contacts prediction methods.

Results:

Protein inter-residue contacts prediction methods are roughly classified into five categories: correlated mutations methods, machine-learning methods, fusion methods, templatebased methods and 3D model-based methods. In this paper, firstly we describe the common definition of protein inter-residue contacts and show the typical application of protein inter-residue contacts. Then, we present a comprehensive review of the three main categories for protein interresidue contacts prediction: correlated mutations methods, machine-learning methods and fusion methods. Besides, we analyze the constraints for each category. Furthermore, we compare several representative methods on the CASP11 dataset and discuss performances of these methods in detail.

Conclusion:

Correlated mutations methods achieve better performances for long-range contacts, while the machine-learning method performs well for short-range contacts. Fusion methods could take advantage of the machine-learning and correlated mutations methods. Employing more effective fusion strategy could be helpful to further improve the performances of fusion methods.

A large-scale comparative assessment of methods for residue-residue contact prediction

Sequence-based prediction of residue-residue contact in proteins becomes increasingly more important for improving protein structure prediction in the big data era. In this study, we performed a large-scale comparative assessment of 15 locally installed contact predictors. To assess these methods, we collected a big data set consisting of 680 nonredundant proteins covering different structural classes and target difficulties. We investigated a wide range of factors that may influence the precision of contact prediction, including target difficulty, structural class, the alignment depth and distribution of contact pairs in a protein structure. We found that: (1) the machine learning-based methods outperform the direct-coupling-based methods for short-range contact prediction, while the latter are significantly better for long-range contact prediction. The consensus-based methods, which combine machine learning and direct-coupling methods, perform the best. (2) The target difficulty does not have clear influence on the machine learning-based methods, while it does affect the direct-coupling and consensus-based methods significantly. (3) The alignment depth has relatively weak effect on the machine learning-based methods. However, for the direct-coupling-based methods and consensus-based methods, the predicted contacts for targets with deeper alignment tend to be more accurate. (4) All methods perform relatively better on β and α + β proteins than on α proteins. (5) Residues buried in the core of protein structure are more prone to be in contact than residues on the surface (22 versus 6%). We believe these are useful results for guiding future development of new approach to contact prediction.

Deep Learning in Biomedical Data Science

Since the 1980s, deep learning and biomedical data have been coevolving and feeding each other. The breadth, complexity, and rapidly expanding size of biomedical data have stimulated the development of novel deep learning methods, and application of these methods to biomedical data have led to scientific discoveries and practical solutions. This overview provides technical and historical pointers to the field, and surveys current applications of deep learning to biomedical data organized around five subareas, roughly of increasing spatial scale: chemoinformatics, proteomics, genomics and transcriptomics, biomedical imaging, and health care. The black box problem of deep learning methods is also briefly discussed.

Two-domain analysis of JmjN-JmjC and PHD-JmjC lysine demethylases: Detecting an inter-domain evolutionary stress

Residues at different positions of a multiple sequence alignment sometimes evolve together, due to a correlated structural or functional stress at these positions. Co-evolution has thus been evidenced computationally in multiple proteins or protein domains. Here, we wish to study whether an evolutionary stress is exerted on a sequence alignment across protein domains, i.e., on longer sequence separations than within a single protein domain. JmjC-containing lysine demethylases were chosen for analysis, as a follow-up to previous studies; these proteins are important multidomain epigenetic regulators. In these proteins, the JmjC domain is responsible for the demethylase activity, and surrounding domains interact with histones, DNA or partner proteins. This family of enzymes was analyzed at the sequence level, in order to determine whether the sequence of JmjC-domains was affected by the presence of a neighboring JmjN domain or PHD finger in the protein. Multiple positions within JmjC sequences were shown to have their residue distributions significantly altered by the presence of the second domain. Structural considerations confirmed the relevance of the analysis for JmjN-JmjC proteins, while among PHD-JmjC proteins, the length of the linker region could be correlated to the residues observed at the most affected positions. The correlation of domain architecture with residue types at certain positions, as well as that of overall architecture with protein function, is discussed. The present results thus evidence the existence of an across-domain evolutionary stress in JmjC-containing demethylases, and provide further insights into the overall domain architecture of JmjC domain-containing proteins.

Co-evolution techniques are reshaping the way we do structural bioinformatics

Co-evolution techniques were originally conceived to assist in protein structure prediction by inferring pairs of residues that share spatial proximity. However, the functional relationships that can be extrapolated from co-evolution have also proven to be useful in a wide array of structural bioinformatics applications. These techniques are a powerful way to extract structural and functional information in a sequence-rich world.

CONFOLD: Residue-residue contact-guided ab initio protein folding

Predicted protein residue-residue contacts can be used to build three-dimensional models and consequently to predict protein folds from scratch. A considerable amount of effort is currently being spent to improve contact prediction accuracy, whereas few methods are available to construct protein tertiary structures from predicted contacts. Here, we present an ab initio protein folding method to build three-dimensional models using predicted contacts and secondary structures. Our method first translates contacts and secondary structures into distance, dihedral angle, and hydrogen bond restraints according to a set of new conversion rules, and then provides these restraints as input for a distance geometry algorithm to build tertiary structure models. The initially reconstructed models are used to regenerate a set of physically realistic contact restraints and detect secondary structure patterns, which are then used to reconstruct final structural models. This unique two-stage modeling approach of integrating contacts and secondary structures improves the quality and accuracy of structural models and in particular generates better β-sheets than other algorithms. We validate our method on two standard benchmark datasets using true contacts and secondary structures. Our method improves TM-score of reconstructed protein models by 45% and 42% over the existing method on the two datasets, respectively. On the dataset for benchmarking reconstructions methods with predicted contacts and secondary structures, the average TM-score of best models reconstructed by our method is 0.59, 5.5% higher than the existing method. The CONFOLD web server is available at http://protein.rnet.missouri.edu/confold/.

Quality assessment of modeled protein structure using physicochemical properties

Physicochemical properties of proteins always guide to determine the quality of the protein structure, therefore it has been rigorously used to distinguish native or native-like structure from other predicted structures. In this work, we explore nine machine learning methods with six physicochemical properties to predict the Root Mean Square Deviation (RMSD), Template Modeling (TM-score), and Global Distance Test (GDT_TS-score) of modeled protein structure in the absence of its true native state. Physicochemical properties namely total surface area, euclidean distance (ED), total empirical energy, secondary structure penalty (SS), sequence length (SL), and pair number (PN) are used. There are a total of 95,091 modeled structures of 4896 native targets. A real coded Self-adaptive Differential Evolution algorithm (SaDE) is used to determine the feature importance. The K-fold cross validation is used to measure the robustness of the best predictive method. Through the intensive experiments, it is found that Random Forest method outperforms over other machine learning methods. This work makes the prediction faster and inexpensive. The performance result shows the prediction of RMSD, TM-score, and GDT_TS-score on Root Mean Square Error (RMSE) as 1.20, 0.06, and 0.06 respectively; correlation scores are 0.96, 0.92, and 0.91 respectively; R(2) are 0.92, 0.85, and 0.84 respectively; and accuracy are 78.82% (with ± 0.1 err), 86.56% (with ± 0.1 err), and 87.37% (with ± 0.1 err) respectively on the testing data set. The data set used in the study is available as supplement at http://bit.ly/RF-PCP-DataSets.

Improved contact predictions using the recognition of protein like contact patterns

Given sufficient large protein families, and using a global statistical inference approach, it is possible to obtain sufficient accuracy in protein residue contact predictions to predict the structure of many proteins. However, these approaches do not consider the fact that the contacts in a protein are neither randomly, nor independently distributed, but actually follow precise rules governed by the structure of the protein and thus are interdependent. Here, we present PconsC2, a novel method that uses a deep learning approach to identify protein-like contact patterns to improve contact predictions. A substantial enhancement can be seen for all contacts independently on the number of aligned sequences, residue separation or secondary structure type, but is largest for β-sheet containing proteins. In addition to being superior to earlier methods based on statistical inferences, in comparison to state of the art methods using machine learning, PconsC2 is superior for families with more than 100 effective sequence homologs. The improved contact prediction enables improved structure prediction.

Here, we introduce a novel protein contact prediction method PconsC2 that, to the best of our knowledge, outperforms earlier methods. PconsC2 is based on our earlier method, PconsC, as it utilizes the same set of contact predictions from plmDCA and PSICOV. However, in contrast to PconsC, where each residue pair is analysed independently, the initial predictions are analysed in context of neighbouring residue pairs using a deep learning approach, inspired by earlier work. We find that for each layer the deep learning procedure improves the predictions. At the end, after five layers of deep learning and inclusion of a few extra features provides the best performance. An improvement can be seen for all types of proteins, independent on length, number of homologous sequences and structural class. However, the improvement is largest for β-sheet containing proteins. Most importantly the improvement brings for the first time sufficiently accurate predictions to some protein families with less than 1000 homologous sequences. PconsC2 outperforms as well state of the art machine learning based predictors for protein families larger than 100 effective sequences. PconsC2 is licensed under the GNU General Public License v3 and freely available from http://c2.pcons.net/.

De novo structure prediction of globular proteins aided by sequence variation-derived contacts

The advent of high accuracy residue-residue intra-protein contact prediction methods enabled a significant boost in the quality of de novo structure predictions. Here, we investigate the potential benefits of combining a well-established fragment-based folding algorithm--FRAGFOLD, with PSICOV, a contact prediction method which uses sparse inverse covariance estimation to identify co-varying sites in multiple sequence alignments. Using a comprehensive set of 150 diverse globular target proteins, up to 266 amino acids in length, we are able to address the effectiveness and some limitations of such approaches to globular proteins in practice. Overall we find that using fragment assembly with both statistical potentials and predicted contacts is significantly better than either statistical potentials or contacts alone. Results show up to nearly 80% of correct predictions (TM-score ≥0.5) within analysed dataset and a mean TM-score of 0.54. Unsuccessful modelling cases emerged either from conformational sampling problems, or insufficient contact prediction accuracy. Nevertheless, a strong dependency of the quality of final models on the fraction of satisfied predicted long-range contacts was observed. This not only highlights the importance of these contacts on determining the protein fold, but also (combined with other ensemble-derived qualities) provides a powerful guide as to the choice of correct models and the global quality of the selected model. A proposed quality assessment scoring function achieves 0.93 precision and 0.77 recall for the discrimination of correct folds on our dataset of decoys. These findings suggest the approach is well-suited for blind predictions on a variety of globular proteins of unknown 3D structure, provided that enough homologous sequences are available to construct a large and accurate multiple sequence alignment for the initial contact prediction step.

Direct coupling analysis for protein contact prediction

During evolution, structure, and function of proteins are remarkably conserved, whereas amino-acid sequences vary strongly between homologous proteins. Structural conservation constrains sequence variability and forces different residues to coevolve, i.e., to show correlated patterns of amino-acid occurrences. However, residue correlation may result from direct coupling, e.g., by a contact in the folded protein, or be induced indirectly via intermediate residues. To use empirically observed correlations for predicting residue-residue contacts, direct and indirect effects have to be disentangled. Here we present mechanistic details on how to achieve this using a methodology called Direct Coupling Analysis (DCA). DCA has been shown to produce highly accurate estimates of amino-acid pairs that have direct reciprocal constraints in evolution. Specifically, we provide instructions and protocols on how to use the algorithmic implementations of DCA starting from data extraction to predicted-contact visualization in contact maps or representative protein structures.

Probabilistic grammatical model for helix-helix contact site classification

Background

Hidden Markov Models power many state‐of‐the‐art tools in the field of protein bioinformatics. While excelling in their tasks, these methods of protein analysis do not convey directly information on medium‐ and long‐range residue‐residue interactions. This requires an expressive power of at least context‐free grammars. However, application of more powerful grammar formalisms to protein analysis has been surprisingly limited.

Results

In this work, we present a probabilistic grammatical framework for problem‐specific protein languages and apply it to classification of transmembrane helix‐helix pairs configurations. The core of the model consists of a probabilistic context‐free grammar, automatically inferred by a genetic algorithm from only a generic set of expert‐based rules and positive training samples. The model was applied to produce sequence based descriptors of four classes of transmembrane helix‐helix contact site configurations. The highest performance of the classifiers reached AUCROC of 0.70. The analysis of grammar parse trees revealed the ability of representing structural features of helix‐helix contact sites.

Conclusions

We demonstrated that our probabilistic context‐free framework for analysis of protein sequences outperforms the state of the art in the task of helix‐helix contact site classification. However, this is achieved without necessarily requiring modeling long range dependencies between interacting residues. A significant feature of our approach is that grammar rules and parse trees are human‐readable. Thus they could provide biologically meaningful information for molecular biologists.

A study and benchmark of DNcon: a method for protein residue-residue contact prediction using deep networks

BACKGROUND

In recent years, the use and importance of predicted protein residue-residue contacts has grown considerably with demonstrated applications such as drug design, protein tertiary structure prediction and model quality assessment. Nevertheless, reported accuracies in the range of 25-35% stubbornly remain the norm for sequence based, long range contact predictions on hard targets. This is in spite of a prolonged effort on behalf of the community to improve the performance of residue-residue contact prediction. A thorough study of the quality of current residue-residue contact predictions and the evaluation metrics used as well as an analysis of current methods is needed to stimulate further advancement in contact prediction and its application. Such a study will better explain the quality and nature of residue-residue contact predictions generated by current methods and as a result lead to better use of this contact information.

RESULTS

We evaluated several sequence based residue-residue contact predictors that participated in the tenth Critical Assessment of protein Structure Prediction (CASP) experiment. The evaluation was performed using standard assessment techniques such as those used by the official CASP assessors as well as two novel evaluation metrics (i.e., cluster accuracy and cluster count). An in-depth analysis revealed that while most residue-residue contact predictions generated are not accurate at the residue level, there is quite a strong contact signal present when allowing for less than residue level precision. Our residue-residue contact predictor, DNcon, performed particularly well achieving an accuracy of 66% for the top L/10 long range contacts when evaluated in a neighbourhood of size 2. The coverage of residue-residue contact areas was also greater with DNcon when compared to other methods. We also provide an analysis of DNcon with respect to its underlying architecture and features used for classification.

CONCLUSIONS

Our novel evaluation metrics demonstrate that current residue-residue contact predictions do contain a strong contact signal and are of better quality than standard evaluation metrics indicate. Our method, DNcon, is a robust, state-of-the-art residue-residue sequence based contact predictor and excelled under a number of evaluation schemes. It is available as a web service at http://iris.rnet.missouri.edu/dncon/.

Evaluation of residue-residue contact prediction in CASP10

We present the results of the assessment of the intramolecular residue-residue contact predictions from 26 prediction groups participating in the 10th round of the CASP experiment. The most recently developed direct coupling analysis methods did not take part in the experiment likely because they require a very deep sequence alignment not available for any of the 114 CASP10 targets. The performance of contact prediction methods was evaluated with the measures used in previous CASPs (i.e., prediction accuracy and the difference between the distribution of the predicted contacts and that of all pairs of residues in the target protein), as well as new measures, such as the Matthews correlation coefficient, the area under the precision-recall curve and the ranks of the first correctly and incorrectly predicted contact. We also evaluated the ability to detect interdomain contacts and tested whether the difficulty of predicting contacts depends upon the protein length and the depth of the family sequence alignment. The analyses were carried out on the target domains for which structural homologs did not exist or were difficult to identify. The evaluation was performed for all types of contacts (short, medium, and long-range), with emphasis placed on long-range contacts, i.e. those involving residues separated by at least 24 residues along the sequence. The assessment suggests that the best CASP10 contact prediction methods perform at approximately the same level, and comparably to those participating in CASP9.

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.

Prediction of contact residue pairs based on co-substitution between sites in protein structures

Residue-residue interactions that fold a protein into a unique three-dimensional structure and make it play a specific function impose structural and functional constraints in varying degrees on each residue site. Selective constraints on residue sites are recorded in amino acid orders in homologous sequences and also in the evolutionary trace of amino acid substitutions. A challenge is to extract direct dependences between residue sites by removing phylogenetic correlations and indirect dependences through other residues within a protein or even through other molecules. Rapid growth of protein families with unknown folds requires an accurate de novo prediction method for protein structure. Recent attempts of disentangling direct from indirect dependences of amino acid types between residue positions in multiple sequence alignments have revealed that inferred residue-residue proximities can be sufficient information to predict a protein fold without the use of known three-dimensional structures. Here, we propose an alternative method of inferring coevolving site pairs from concurrent and compensatory substitutions between sites in each branch of a phylogenetic tree. Substitution probability and physico-chemical changes (volume, charge, hydrogen-bonding capability, and others) accompanied by substitutions at each site in each branch of a phylogenetic tree are estimated with the likelihood of each substitution, and their direct correlations between sites are used to detect concurrent and compensatory substitutions. In order to extract direct dependences between sites, partial correlation coefficients of the characteristic changes along branches between sites, in which linear multiple dependences on feature vectors at other sites are removed, are calculated and used to rank coevolving site pairs. Accuracy of contact prediction based on the present coevolution score is comparable to that achieved by a maximum entropy model of protein sequences for 15 protein families taken from the Pfam release 26.0. Besides, this excellent accuracy indicates that compensatory substitutions are significant in protein evolution.

Effective inter-residue contact definitions for accurate protein fold recognition

BACKGROUND

Effective encoding of residue contact information is crucial for protein structure prediction since it has a unique role to capture long-range residue interactions compared to other commonly used scoring terms. The residue contact information can be incorporated in structure prediction in several different ways: It can be incorporated as statistical potentials or it can be also used as constraints in ab initio structure prediction. To seek the most effective definition of residue contacts for template-based protein structure prediction, we evaluated 45 different contact definitions, varying bases of contacts and distance cutoffs, in terms of their ability to identify proteins of the same fold.

RESULTS

We found that overall the residue contact pattern can distinguish protein folds best when contacts are defined for residue pairs whose Cβ atoms are at 7.0 Å or closer to each other. Lower fold recognition accuracy was observed when inaccurate threading alignments were used to identify common residue contacts between protein pairs. In the case of threading, alignment accuracy strongly influences the fraction of common contacts identified among proteins of the same fold, which eventually affects the fold recognition accuracy. The largest deterioration of the fold recognition was observed for β-class proteins when the threading methods were used because the average alignment accuracy was worst for this fold class. When results of fold recognition were examined for individual proteins, we found that the effective contact definition depends on the fold of the proteins. A larger distance cutoff is often advantageous for capturing spatial arrangement of the secondary structures which are not physically in contact. For capturing contacts between neighboring β strands, considering the distance between Cα atoms is better than the Cβ-based distance because the side-chain of interacting residues on β strands sometimes point to opposite directions.

CONCLUSION

Residue contacts defined by Cβ-Cβ distance of 7.0 Å work best overall among tested to identify proteins of the same fold. We also found that effective contact definitions differ from fold to fold, suggesting that using different residue contact definition specific for each template will lead to improvement of the performance of threading.

Predicting protein residue-residue contacts using deep networks and boosting

MOTIVATION

Protein residue-residue contacts continue to play a larger and larger role in protein tertiary structure modeling and evaluation. Yet, while the importance of contact information increases, the performance of sequence-based contact predictors has improved slowly. New approaches and methods are needed to spur further development and progress in the field.

RESULTS

Here we present DNCON, a new sequence-based residue-residue contact predictor using deep networks and boosting techniques. Making use of graphical processing units and CUDA parallel computing technology, we are able to train large boosted ensembles of residue-residue contact predictors achieving state-of-the-art performance.

AVAILABILITY

The web server of the prediction method (DNCON) is available at http://iris.rnet.missouri.edu/dncon/.

CONTACT

chengji@missouri.edu

SUPPLEMENTARY INFORMATION

Supplementary data are available at Bioinformatics online.

Genomics-aided structure prediction

We introduce a theoretical framework that exploits the ever-increasing genomic sequence information for protein structure prediction. Structure-based models are modified to incorporate constraints by a large number of non-local contacts estimated from direct coupling analysis (DCA) of co-evolving genomic sequences. A simple hybrid method, called DCA-fold, integrating DCA contacts with an accurate knowledge of local information (e.g., the local secondary structure) is sufficient to fold proteins in the range of 1-3 Å resolution.

Accurate de novo structure prediction of large transmembrane protein domains using fragment-assembly and correlated mutation analysis

A new de novo protein structure prediction method for transmembrane proteins (FILM3) is described that is able to accurately predict the structures of large membrane proteins domains using an ensemble of two secondary structure prediction methods to guide fragment selection in combination with a scoring function based solely on correlated mutations detected in multiple sequence alignments. This approach has been validated by generating models for 28 membrane proteins with a diverse range of complex topologies and an average length of over 300 residues with results showing that TM-scores > 0.5 can be achieved in almost every case following refinement using MODELLER. In one of the most impressive results, a model of mitochondrial cytochrome c oxidase polypeptide I was obtained with a TM-score > 0.75 and an rmsd of only 5.7 Å over all 514 residues. These results suggest that FILM3 could be applicable to a wide range of transmembrane proteins of as-yet-unknown 3D structure given sufficient homologous sequences.

Protein 3D structure computed from evolutionary sequence variation

The evolutionary trajectory of a protein through sequence space is constrained by its function. Collections of sequence homologs record the outcomes of millions of evolutionary experiments in which the protein evolves according to these constraints. Deciphering the evolutionary record held in these sequences and exploiting it for predictive and engineering purposes presents a formidable challenge. The potential benefit of solving this challenge is amplified by the advent of inexpensive high-throughput genomic sequencing.

In this paper we ask whether we can infer evolutionary constraints from a set of sequence homologs of a protein. The challenge is to distinguish true co-evolution couplings from the noisy set of observed correlations. We address this challenge using a maximum entropy model of the protein sequence, constrained by the statistics of the multiple sequence alignment, to infer residue pair couplings. Surprisingly, we find that the strength of these inferred couplings is an excellent predictor of residue-residue proximity in folded structures. Indeed, the top-scoring residue couplings are sufficiently accurate and well-distributed to define the 3D protein fold with remarkable accuracy.

We quantify this observation by computing, from sequence alone, all-atom 3D structures of fifteen test proteins from different fold classes, ranging in size from 50 to 260 residues., including a G-protein coupled receptor. These blinded inferences are de novo, i.e., they do not use homology modeling or sequence-similar fragments from known structures. The co-evolution signals provide sufficient information to determine accurate 3D protein structure to 2.7–4.8 Å C_α-RMSD error relative to the observed structure, over at least two-thirds of the protein (method called EVfold, details at http://EVfold.org). This discovery provides insight into essential interactions constraining protein evolution and will facilitate a comprehensive survey of the universe of protein structures, new strategies in protein and drug design, and the identification of functional genetic variants in normal and disease genomes.

PREDICTION OF CONTACT MAPS USING SUPPORT VECTOR MACHINES

Contact map prediction is of great interest for its application in fold recognition and protein 3D structure determination. In this paper we present a contact-map prediction algorithm that employs Support Vector Machines as the machine learning tool and incorporates various features such as sequence profiles and their conservations, correlated mutation analysis based on various amino acid physicochemical properties, and secondary structure. In addition, we evaluated the effectiveness of the different features on contact map prediction for different fold classes. On average, our predictor achieved a prediction accuracy of 0.224 with an improvement over a random predictor of a factor 11.7, which is better than reported studies. Our study showed that predicted secondary structure features play an important roles for the proteins containing beta-structures. Models based on secondary structure features and correlated mutation analysis features produce different sets of predictions. Our study also suggests that models learned separately for different protein fold families may achieve better performance than a unified model.

PSICOV: precise structural contact prediction using sparse inverse covariance estimation on large multiple sequence alignments

MOTIVATION

The accurate prediction of residue-residue contacts, critical for maintaining the native fold of a protein, remains an open problem in the field of structural bioinformatics. Interest in this long-standing problem has increased recently with algorithmic improvements and the rapid growth in the sizes of sequence families. Progress could have major impacts in both structure and function prediction to name but two benefits. Sequence-based contact predictions are usually made by identifying correlated mutations within multiple sequence alignments (MSAs), most commonly through the information-theoretic approach of calculating mutual information between pairs of sites in proteins. These predictions are often inaccurate because the true covariation signal in the MSA is often masked by biases from many ancillary indirect-coupling or phylogenetic effects. Here we present a novel method, PSICOV, which introduces the use of sparse inverse covariance estimation to the problem of protein contact prediction. Our method builds on work which had previously demonstrated corrections for phylogenetic and entropic correlation noise and allows accurate discrimination of direct from indirectly coupled mutation correlations in the MSA.

RESULTS

PSICOV displays a mean precision substantially better than the best performing normalized mutual information approach and Bayesian networks. For 118 out of 150 targets, the L/5 (i.e. top-L/5 predictions for a protein of length L) precision for long-range contacts (sequence separation >23) was ≥ 0.5, which represents an improvement sufficient to be of significant benefit in protein structure prediction or model quality assessment.

AVAILABILITY

The PSICOV source code can be downloaded from http://bioinf.cs.ucl.ac.uk/downloads/PSICOV.

Predicting residue-residue contacts and helix-helix interactions in transmembrane proteins using an integrative feature-based random forest approach

Integral membrane proteins constitute 25-30% of genomes and play crucial roles in many biological processes. However, less than 1% of membrane protein structures are in the Protein Data Bank. In this context, it is important to develop reliable computational methods for predicting the structures of membrane proteins. Here, we present the first application of random forest (RF) for residue-residue contact prediction in transmembrane proteins, which we term as TMhhcp. Rigorous cross-validation tests indicate that the built RF models provide a more favorable prediction performance compared with two state-of-the-art methods, i.e., TMHcon and MEMPACK. Using a strict leave-one-protein-out jackknifing procedure, they were capable of reaching the top L/5 prediction accuracies of 49.5% and 48.8% for two different residue contact definitions, respectively. The predicted residue contacts were further employed to predict interacting helical pairs and achieved the Matthew's correlation coefficients of 0.430 and 0.424, according to two different residue contact definitions, respectively. To facilitate the academic community, the TMhhcp server has been made freely accessible at http://protein.cau.edu.cn/tmhhcp.

Enhanced Inter-helical Residue Contact Prediction in Transmembrane Proteins

In this paper, based on a recent work by McAllister and Floudas who developed a mathematical optimization model to predict the contacts in transmembrane alpha-helical proteins from a limited protein data set [1], we have enhanced this method by 1) building a more comprehensive data set for transmembrane alpha-helical proteins and this enhanced data set is then used to construct the probability sets, MIN-1N and MIN-2N, for residue contact prediction, 2) enhancing the mathematical model via modifications of several important physical constraints and 3) applying a new blind contact prediction scheme on different protein sets proposed from analyzing the contact prediction on 65 proteins from Fuchs et al. [2]. The blind contact prediction scheme has been tested on two different membrane protein sets. Firstly it is applied to five carefully selected proteins from the training set. The contact prediction of these five proteins uses probability sets built by excluding the target protein from the training set, and an average accuracy of 56% was obtained. Secondly, it is applied to six independent membrane proteins with complicated topologies, and the prediction accuracies are 73% for 2ZY9A, 21% for 3KCUA, 46% for 2W1PA, 64% for 3CN5A, 77% for 3IXZA and 83% for 3K3FA. The average prediction accuracy for the six proteins is 60.7%. The proposed approach is also compared with a support vector machine method (TMhit [3]) and it is shown that it exhibits better prediction accuracy.

Evaluation of residue-residue contact predictions in CASP9

This work presents the results of the assessment of the intramolecular residue-residue contact predictions submitted to CASP9. The methodology for the assessment does not differ from that used in previous CASPs, with two basic evaluation measures being the precision in recognizing contacts and the difference between the distribution of distances in the subset of predicted contact pairs versus all pairs of residues in the structure. The emphasis is placed on the prediction of long-range contacts (i.e., contacts between residues separated by at least 24 residues along sequence) in target proteins that cannot be easily modeled by homology. Although there is considerable activity in the field, the current analysis reports no discernable progress since CASP8.

Correlated mutations: a hallmark of phenotypic amino acid substitutions

Point mutations resulting in the substitution of a single amino acid can cause severe functional consequences, but can also be completely harmless. Understanding what determines the phenotypical impact is important both for planning targeted mutation experiments in the laboratory and for analyzing naturally occurring mutations found in patients. Common wisdom suggests using the extent of evolutionary conservation of a residue or a sequence motif as an indicator of its functional importance and thus vulnerability in case of mutation. In this work, we put forward the hypothesis that in addition to conservation, co-evolution of residues in a protein influences the likelihood of a residue to be functionally important and thus associated with disease. While the basic idea of a relation between co-evolution and functional sites has been explored before, we have conducted the first systematic and comprehensive analysis of point mutations causing disease in humans with respect to correlated mutations. We included 14,211 distinct positions with known disease-causing point mutations in 1,153 human proteins in our analysis. Our data show that (1) correlated positions are significantly more likely to be disease-associated than expected by chance, and that (2) this signal cannot be explained by conservation patterns of individual sequence positions. Although correlated residues have primarily been used to predict contact sites, our data are in agreement with previous observations that (3) many such correlations do not relate to physical contacts between amino acid residues. Access to our analysis results are provided at http://webclu.bio.wzw.tum.de/~pagel/supplements/correlated-positions/.

Point mutations (i.e., changes of a single sequence element) can have a severe impact on protein function. Many diseases are caused by such minute defects. On the other hand, the majority of such mutations does not lead to noticeable effects. Although previous research has revealed important aspects that influence or predict the chance of a mutation to cause disease, much remains to be learned before we fully understand this complex problem. In our work, we use the observation that sometimes certain positions in a protein mutate in an apparently correlated fashion and analyze this correlation with respect to mutation vulnerability. Our results show that positions exhibiting evolutionary correlation are significantly more likely to be vulnerable to mutation than average positions. On one hand, our data further support the concept of correlated positions to not only be associated with protein contacts but also functional sites and/or disease positions (as introduced by others). On the other hand, this could be useful to further improve the understanding and prediction of the consequences of mutations. Our work is the first to attempt a large-scale quantitation of this relationship.