High precision in protein contact prediction using fully convolutional neural networks and minimal sequence features

COMTOP: Protein Residue-Residue Contact Prediction through Mixed Integer Linear Optimization

Protein contact prediction helps reconstruct the tertiary structure that greatly determines a protein’s function; therefore, contact prediction from the sequence is an important problem. Recently there has been exciting progress on this problem, but many of the existing methods are still low quality of prediction accuracy. In this paper, we present a new mixed integer linear programming (MILP)-based consensus method: a Consensus scheme based On a Mixed integer linear opTimization method for prOtein contact Prediction (COMTOP). The MILP-based consensus method combines the strengths of seven selected protein contact prediction methods, including CCMpred, EVfold, DeepCov, NNcon, PconsC4, plmDCA, and PSICOV, by optimizing the number of correctly predicted contacts and achieving a better prediction accuracy. The proposed hybrid protein residue–residue contact prediction scheme was tested in four independent test sets. For 239 highly non-redundant proteins, the method showed a prediction accuracy of 59.68%, 70.79%, 78.86%, 89.04%, 94.51%, and 97.35% for top-5L, top-3L, top-2L, top-L, top-L/2, and top-L/5 contacts, respectively. When tested on the CASP13 and CASP14 test sets, the proposed method obtained accuracies of 75.91% and 77.49% for top-L/5 predictions, respectively. COMTOP was further tested on 57 non-redundant α-helical transmembrane proteins and achieved prediction accuracies of 64.34% and 73.91% for top-L/2 and top-L/5 predictions, respectively. For all test datasets, the improvement of COMTOP in accuracy over the seven individual methods increased with the increasing number of predicted contacts. For example, COMTOP performed much better for large number of contact predictions (such as top-5L and top-3L) than for small number of contact predictions such as top-L/2 and top-L/5. The results and analysis demonstrate that COMTOP can significantly improve the performance of the individual methods; therefore, COMTOP is more robust against different types of test sets. COMTOP also showed better/comparable predictions when compared with the state-of-the-art predictors.

DNCON2_Inter: predicting interchain contacts for homodimeric and homomultimeric protein complexes using multiple sequence alignments of monomers and deep learning

Deep learning methods that achieved great success in predicting intrachain residue-residue contacts have been applied to predict interchain contacts between proteins. However, these methods require multiple sequence alignments (MSAs) of a pair of interacting proteins (dimers) as input, which are often difficult to obtain because there are not many known protein complexes available to generate MSAs of sufficient depth for a pair of proteins. In recognizing that multiple sequence alignments of a monomer that forms homomultimers contain the co-evolutionary signals of both intrachain and interchain residue pairs in contact, we applied DNCON2 (a deep learning-based protein intrachain residue-residue contact predictor) to predict both intrachain and interchain contacts for homomultimers using multiple sequence alignment (MSA) and other co-evolutionary features of a single monomer followed by discrimination of interchain and intrachain contacts according to the tertiary structure of the monomer. We name this tool DNCON2_Inter. Allowing true-positive predictions within two residue shifts, the best average precision was obtained for the Top-L/10 predictions of 22.9% for homodimers and 17.0% for higher-order homomultimers. In some instances, especially where interchain contact densities are high, DNCON2_Inter predicted interchain contacts with 100% precision. We also developed Con_Complex, a complex structure reconstruction tool that uses predicted contacts to produce the structure of the complex. Using Con_Complex, we show that the predicted contacts can be used to accurately construct the structure of some complexes. Our experiment demonstrates that monomeric multiple sequence alignments can be used with deep learning to predict interchain contacts of homomeric proteins.

Recent Applications of Deep Learning Methods on Evolution- and Contact-Based Protein Structure Prediction

The new advances in deep learning methods have influenced many aspects of scientific research, including the study of the protein system. The prediction of proteins' 3D structural components is now heavily dependent on machine learning techniques that interpret how protein sequences and their homology govern the inter-residue contacts and structural organization. Especially, methods employing deep neural networks have had a significant impact on recent CASP13 and CASP14 competition. Here, we explore the recent applications of deep learning methods in the protein structure prediction area. We also look at the potential opportunities for deep learning methods to identify unknown protein structures and functions to be discovered and help guide drug-target interactions. Although significant problems still need to be addressed, we expect these techniques in the near future to play crucial roles in protein structural bioinformatics as well as in drug discovery.

Protein Structure Prediction: Conventional and Deep Learning Perspectives

Protein structure prediction is a way to bridge the sequence-structure gap, one of the main challenges in computational biology and chemistry. Predicting any protein's accurate structure is of paramount importance for the scientific community, as these structures govern their function. Moreover, this is one of the complicated optimization problems that computational biologists have ever faced. Experimental protein structure determination methods include X-ray crystallography, Nuclear Magnetic Resonance Spectroscopy and Electron Microscopy. All of these are tedious and time-consuming procedures that require expertise. To make the process less cumbersome, scientists use predictive tools as part of computational methods, using data consolidated in the protein repositories. In recent years, machine learning approaches have raised the interest of the structure prediction community. Most of the machine learning approaches for protein structure prediction are centred on co-evolution based methods. The accuracy of these approaches depends on the number of homologous protein sequences available in the databases. The prediction problem becomes challenging for many proteins, especially those without enough sequence homologs. Deep learning methods allow for the extraction of intricate features from protein sequence data without making any intuitions. Accurately predicted protein structures are employed for drug discovery, antibody designs, understanding protein-protein interactions, and interactions with other molecules. This article provides a review of conventional and deep learning approaches in protein structure prediction. We conclude this review by outlining a few publicly available datasets and deep learning architectures currently employed for protein structure prediction tasks.

On the effect of phylogenetic correlations in coevolution-based contact prediction in proteins

Coevolution-based contact prediction, either directly by coevolutionary couplings resulting from global statistical sequence models or using structural supervision and deep learning, has found widespread application in protein-structure prediction from sequence. However, one of the basic assumptions in global statistical modeling is that sequences form an at least approximately independent sample of an unknown probability distribution, which is to be learned from data. In the case of protein families, this assumption is obviously violated by phylogenetic relations between protein sequences. It has turned out to be notoriously difficult to take phylogenetic correlations into account in coevolutionary model learning. Here, we propose a complementary approach: we develop strategies to randomize or resample sequence data, such that conservation patterns and phylogenetic relations are preserved, while intrinsic (i.e. structure- or function-based) coevolutionary couplings are removed. A comparison between the results of Direct Coupling Analysis applied to real and to resampled data shows that the largest coevolutionary couplings, i.e. those used for contact prediction, are only weakly influenced by phylogeny. However, the phylogeny-induced spurious couplings in the resampled data are compatible in size with the first false-positive contact predictions from real data. Dissecting functional from phylogeny-induced couplings might therefore extend accurate contact predictions to the range of intermediate-size couplings.

Many homologous protein families contain thousands of highly diverged amino-acid sequences, which fold into close-to-identical three-dimensional structures and fulfill almost identical biological tasks. Global coevolutionary models, like those inferred by the Direct Coupling Analysis (DCA), assume that families can be considered as samples of some unknown statistical model, and that the parameters of these models represent evolutionary constraints acting on protein sequences. To learn these models from data, DCA and related approaches have to also assume that the distinct sequences in a protein family are close to independent, while in reality they are characterized by involved hierarchical phylogenetic relationships. Here we propose Null models for sequence alignments, which maintain patterns of amino-acid conservation and phylogeny contained in the data, but destroy any coevolutionary couplings, frequently used in protein structure prediction. We find that phylogeny actually induces spurious non-zero couplings. These are, however, significantly smaller that the largest couplings derived from natural sequences, and therefore have only little influence on the first predicted contacts. However, in the range of intermediate couplings, they may lead to statistically significant effects. Dissecting phylogenetic from functional couplings might therefore extend the range of accurately predicted structural contacts down to smaller coupling strengths than those currently used.

Deep Learning-Based Advances in Protein Structure Prediction

Obtaining an accurate description of protein structure is a fundamental step toward understanding the underpinning of biology. Although recent advances in experimental approaches have greatly enhanced our capabilities to experimentally determine protein structures, the gap between the number of protein sequences and known protein structures is ever increasing. Computational protein structure prediction is one of the ways to fill this gap. Recently, the protein structure prediction field has witnessed a lot of advances due to Deep Learning (DL)-based approaches as evidenced by the success of AlphaFold2 in the most recent Critical Assessment of protein Structure Prediction (CASP14). In this article, we highlight important milestones and progresses in the field of protein structure prediction due to DL-based methods as observed in CASP experiments. We describe advances in various steps of protein structure prediction pipeline viz. protein contact map prediction, protein distogram prediction, protein real-valued distance prediction, and Quality Assessment/refinement. We also highlight some end-to-end DL-based approaches for protein structure prediction approaches. Additionally, as there have been some recent DL-based advances in protein structure determination using Cryo-Electron (Cryo-EM) microscopy based, we also highlight some of the important progress in the field. Finally, we provide an outlook and possible future research directions for DL-based approaches in the protein structure prediction arena.

Evaluation of residue-residue contact prediction methods: From retrospective to prospective

Sequence-based residue contact prediction plays a crucial role in protein structure reconstruction. In recent years, the combination of evolutionary coupling analysis (ECA) and deep learning (DL) techniques has made tremendous progress for residue contact prediction, thus a comprehensive assessment of current methods based on a large-scale benchmark data set is very needed. In this study, we evaluate 18 contact predictors on 610 non-redundant proteins and 32 CASP13 targets according to a wide range of perspectives. The results show that different methods have different application scenarios: (1) DL methods based on multi-categories of inputs and large training sets are the best choices for low-contact-density proteins such as the intrinsically disordered ones and proteins with shallow multi-sequence alignments (MSAs). (2) With at least 5L (L is sequence length) effective sequences in the MSA, all the methods show the best performance, and methods that rely only on MSA as input can reach comparable achievements as methods that adopt multi-source inputs. (3) For top L/5 and L/2 predictions, DL methods can predict more hydrophobic interactions while ECA methods predict more salt bridges and disulfide bonds. (4) ECA methods can detect more secondary structure interactions, while DL methods can accurately excavate more contact patterns and prune isolated false positives. In general, multi-input DL methods with large training sets dominate current approaches with the best overall performance. Despite the great success of current DL methods must be stated the fact that there is still much room left for further improvement: (1) With shallow MSAs, the performance will be greatly affected. (2) Current methods show lower precisions for inter-domain compared with intra-domain contact predictions, as well as very high imbalances in precisions between intra-domains. (3) Strong prediction similarities between DL methods indicating more feature types and diversified models need to be developed. (4) The runtime of most methods can be further optimized.

The amino acid sequence of a protein ultimately determines its tertiary structure, and the tertiary structure determines its function(s) and plays a key role in understanding biological processes and disease pathogenesis. Protein tertiary structure can be determined using experimental techniques such as cryo-electron microscopy, nuclear magnetic resonance and X-ray crystallography, which are very expensive and time-consuming. As an alternative, researchers are trying to use in silico methods to predict the 3D structures. Residue contact-assisted protein folding paves an avenue for sequence-based protein structure prediction and therefore has become one of the most challenging and promising problems in structural bioinformatics. Over the past years, contact prediction has undergone continuous evolution in techniques. Through a retrospective analysis of traditional machine learning /evolutionary coupling analysis methods/ consensus machine learning methods and a multi-perspective study on recently developed deep learning methods, we explore the most advanced contact predictors, pursue application scenarios for different methods, and seek prospective directions for further improvement. We anticipate that our study will serve as a practical and useful guide for the development of future approaches to contact prediction.

AlphaFold and the amyloid landscape

Protein aggregation is a widespread phenomenon with important implications in many scientific areas. Although amyloid formation is typically considered as detrimental, functional amyloids that perform physiological roles have been identified in all kingdoms of life. Despite their functional and pathological relevance, the structural details of the majority of molecular species involved in the amyloidogenic process remains elusive. Here, we explore the application of AlphaFold, a highly accurate protein structure predictor, in the field of protein aggregation. While we envision a straightforward application of AlphaFold in assisting the design of globular proteins with improved solubility for biomedical and industrial purposes, the use of this algorithm for predicting the structure of aggregated species seems far from trivial. First, in amyloid diseases, the presence of multiple amyloid polymorphs and the heterogeneity of aggregation intermediates challenges the "one sequence, one structure" paradigm, inherent to sequence-based predictions. Second, aberrant aggregation is not the subject of positive selective pressure, precluding the use of evolutionary-based approaches, which are the core of the AlphaFold pipeline. Instead, amyloid polymorphism seems to be constrained by the need for a defined structure-activity relationship in functional amyloids. They may thus provide a starting point for the application of AlphaFold in the amyloid landscape.

Improving deep learning-based protein distance prediction in CASP14

Motivation

Accurate prediction of residue–residue distances is important for protein structure prediction. We developed several protein distance predictors based on a deep learning distance prediction method and blindly tested them in the 14th Critical Assessment of Protein Structure Prediction (CASP14). The prediction method uses deep residual neural networks with the channel-wise attention mechanism to classify the distance between every two residues into multiple distance intervals. The input features for the deep learning method include co-evolutionary features as well as other sequence-based features derived from multiple sequence alignments (MSAs). Three alignment methods are used with multiple protein sequence/profile databases to generate MSAs for input feature generation. Based on different configurations and training strategies of the deep learning method, five MULTICOM distance predictors were created to participate in the CASP14 experiment.

Results

Benchmarked on 37 hard CASP14 domains, the best performing MULTICOM predictor is ranked 5th out of 30 automated CASP14 distance prediction servers in terms of precision of top L/5 long-range contact predictions [i.e. classifying distances between two residues into two categories: in contact (<8 Angstrom) and not in contact otherwise] and performs better than the best CASP13 distance prediction method. The best performing MULTICOM predictor is also ranked 6th among automated server predictors in classifying inter-residue distances into 10 distance intervals defined by CASP14 according to the precision of distance classification. The results show that the quality and depth of MSAs depend on alignment methods and sequence databases and have a significant impact on the accuracy of distance prediction. Using larger training datasets and multiple complementary features improves prediction accuracy. However, the number of effective sequences in MSAs is only a weak indicator of the quality of MSAs and the accuracy of predicted distance maps. In contrast, there is a strong correlation between the accuracy of contact/distance predictions and the average probability of the predicted contacts, which can therefore be more effectively used to estimate the confidence of distance predictions and select predicted distance maps.

Availability and implementation

The software package, source code and data of DeepDist2 are freely available at https://github.com/multicom-toolbox/deepdist and https://zenodo.org/record/4712084#.YIIM13VKhQM.

Supplementary information

Supplementary data are available at Bioinformatics online.

GraphQA: protein model quality assessment using graph convolutional networks

Motivation

Proteins are ubiquitous molecules whose function in biological processes is determined by their 3D structure. Experimental identification of a protein’s structure can be time-consuming, prohibitively expensive and not always possible. Alternatively, protein folding can be modeled using computational methods, which however are not guaranteed to always produce optimal results. GraphQA is a graph-based method to estimate the quality of protein models, that possesses favorable properties such as representation learning, explicit modeling of both sequential and 3D structure, geometric invariance and computational efficiency.

Results

GraphQA performs similarly to state-of-the-art methods despite using a relatively low number of input features. In addition, the graph network structure provides an improvement over the architecture used in ProQ4 operating on the same input features. Finally, the individual contributions of GraphQA components are carefully evaluated.

Availability and implementation

PyTorch implementation, datasets, experiments and link to an evaluation server are available through this GitHub repository: github.com/baldassarreFe/graphqa.

Supplementary information

Supplementary data are available at Bioinformatics online.

Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences

In the field of artificial intelligence, a combination of scale in data and model capacity enabled by unsupervised learning has led to major advances in representation learning and statistical generation. In the life sciences, the anticipated growth of sequencing promises unprecedented data on natural sequence diversity. Protein language modeling at the scale of evolution is a logical step toward predictive and generative artificial intelligence for biology. To this end, we use unsupervised learning to train a deep contextual language model on 86 billion amino acids across 250 million protein sequences spanning evolutionary diversity. The resulting model contains information about biological properties in its representations. The representations are learned from sequence data alone. The learned representation space has a multiscale organization reflecting structure from the level of biochemical properties of amino acids to remote homology of proteins. Information about secondary and tertiary structure is encoded in the representations and can be identified by linear projections. Representation learning produces features that generalize across a range of applications, enabling state-of-the-art supervised prediction of mutational effect and secondary structure and improving state-of-the-art features for long-range contact prediction.

Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences

In the field of artificial intelligence, a combination of scale in data and model capacity enabled by unsupervised learning has led to major advances in representation learning and statistical generation. In the life sciences, the anticipated growth of sequencing promises unprecedented data on natural sequence diversity. Protein language modeling at the scale of evolution is a logical step toward predictive and generative artificial intelligence for biology. To this end, we use unsupervised learning to train a deep contextual language model on 86 billion amino acids across 250 million protein sequences spanning evolutionary diversity. The resulting model contains information about biological properties in its representations. The representations are learned from sequence data alone. The learned representation space has a multiscale organization reflecting structure from the level of biochemical properties of amino acids to remote homology of proteins. Information about secondary and tertiary structure is encoded in the representations and can be identified by linear projections. Representation learning produces features that generalize across a range of applications, enabling state-of-the-art supervised prediction of mutational effect and secondary structure and improving state-of-the-art features for long-range contact prediction.

Crystallographic molecular replacement using an in silico-generated search model of SARS-CoV-2 ORF8

The majority of crystal structures are determined by the method of molecular replacement (MR). The range of application of MR is limited mainly by the need for an accurate search model. In most cases, pre-existing experimentally determined structures are used as search models. In favorable cases, ab initio predicted structures have yielded search models adequate for MR. The ORF8 protein of SARS-CoV-2 represents a challenging case for MR using an ab initio prediction because ORF8 has an all β-sheet fold and few orthologs. We previously determined experimentally the structure of ORF8 using the single anomalous dispersion (SAD) phasing method, having been unable to find an MR solution to the crystallographic phase problem. Following a report of an accurate prediction of the ORF8 structure, we assessed whether the predicted model would have succeeded as an MR search model. A phase problem solution was found, and the resulting structure was refined, yielding structural parameters equivalent to the original experimental solution.

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.

Biological network analysis with deep learning

Recent advancements in experimental high-throughput technologies have expanded the availability and quantity of molecular data in biology. Given the importance of interactions in biological processes, such as the interactions between proteins or the bonds within a chemical compound, this data is often represented in the form of a biological network. The rise of this data has created a need for new computational tools to analyze networks. One major trend in the field is to use deep learning for this goal and, more specifically, to use methods that work with networks, the so-called graph neural networks (GNNs). In this article, we describe biological networks and review the principles and underlying algorithms of GNNs. We then discuss domains in bioinformatics in which graph neural networks are frequently being applied at the moment, such as protein function prediction, protein-protein interaction prediction and in silico drug discovery and development. Finally, we highlight application areas such as gene regulatory networks and disease diagnosis where deep learning is emerging as a new tool to answer classic questions like gene interaction prediction and automatic disease prediction from data.

Biological network analysis with deep learning

Recent advancements in experimental high-throughput technologies have expanded the availability and quantity of molecular data in biology. Given the importance of interactions in biological processes, such as the interactions between proteins or the bonds within a chemical compound, this data is often represented in the form of a biological network. The rise of this data has created a need for new computational tools to analyze networks. One major trend in the field is to use deep learning for this goal and, more specifically, to use methods that work with networks, the so-called graph neural networks (GNNs). In this article, we describe biological networks and review the principles and underlying algorithms of GNNs. We then discuss domains in bioinformatics in which graph neural networks are frequently being applied at the moment, such as protein function prediction, protein-protein interaction prediction and in silico drug discovery and development. Finally, we highlight application areas such as gene regulatory networks and disease diagnosis where deep learning is emerging as a new tool to answer classic questions like gene interaction prediction and automatic disease prediction from data.

Incorporating Machine Learning into Established Bioinformatics Frameworks

The exponential growth of biomedical data in recent years has urged the application of numerous machine learning techniques to address emerging problems in biology and clinical research. By enabling the automatic feature extraction, selection, and generation of predictive models, these methods can be used to efficiently study complex biological systems. Machine learning techniques are frequently integrated with bioinformatic methods, as well as curated databases and biological networks, to enhance training and validation, identify the best interpretable features, and enable feature and model investigation. Here, we review recently developed methods that incorporate machine learning within the same framework with techniques from molecular evolution, protein structure analysis, systems biology, and disease genomics. We outline the challenges posed for machine learning, and, in particular, deep learning in biomedicine, and suggest unique opportunities for machine learning techniques integrated with established bioinformatics approaches to overcome some of these challenges.

Improved sequence-based prediction of interaction sites in α-helical transmembrane proteins by deep learning

•

Fast and accurate prediction of transmembrane protein interaction sites.

•

First ever computational survey of interaction sites in membrane proteins.

•

10-30% of amino acid positions predicted to be involved in interactions.

Interactions between transmembrane (TM) proteins are fundamental for a wide spectrum of cellular functions, but precise molecular details of these interactions remain largely unknown due to the scarcity of experimentally determined three-dimensional complex structures. Computational techniques are therefore required for a large-scale annotation of interaction sites in TM proteins. Here, we present a novel deep-learning approach, DeepTMInter, for sequence-based prediction of interaction sites in α-helical TM proteins based on their topological, physiochemical, and evolutionary properties. Using a combination of ultra-deep residual neural networks with a stacked generalization ensemble technique DeepTMInter significantly outperforms existing methods, achieving the AUC/AUCPR values of 0.689/0.598. Across the main functional families of human transmembrane proteins, the percentage of amino acid sites predicted to be involved in interactions typically ranges between 10% and 25%, and up to 30% in ion channels. DeepTMInter is available as a standalone package at https://github.com/2003100127/deeptminter. The training and benchmarking datasets are available at https://data.mendeley.com/datasets/2t8kgwzp35.

VoroCNN: Deep convolutional neural network built on 3D Voronoi tessellation of protein structures

MOTIVATION

Effective use of evolutionary information has recently led to tremendous progress in computational prediction of three-dimensional (3D) structures of proteins and their complexes. Despite the progress, the accuracy of predicted structures tends to vary considerably from case to case. Since the utility of computational models depends on their accuracy, reliable estimates of deviation between predicted and native structures are of utmost importance.

RESULTS

For the first time, we present a deep convolutional neural network (CNN) constructed on a Voronoi tessellation of 3D molecular structures. Despite the irregular data domain, our data representation allows us to efficiently introduce both convolution and pooling operations and train the network in an end-to-end fashion without precomputed descriptors. The resultant model, VoroCNN, predicts local qualities of 3D protein folds. The prediction results are competitive to state of the art and superior to the previous 3D CNN architectures built for the same task. We also discuss practical applications of VoroCNN, for example, in recognition of protein binding interfaces.

AVAILABILITY

The model, data, and evaluation tests are available at https://team.inria.fr/nano-d/software/vorocnn/.

SUPPLEMENTARY INFORMATION

Supplementary data are available at Bioinformatics online.

Why can deep convolutional neural networks improve protein fold recognition? A visual explanation by interpretation

As an essential task in protein structure and function prediction, protein fold recognition has attracted increasing attention. The majority of the existing machine learning-based protein fold recognition approaches strongly rely on handcrafted features, which depict the characteristics of different protein folds; however, effective feature extraction methods still represent the bottleneck for further performance improvement of protein fold recognition. As a powerful feature extractor, deep convolutional neural network (DCNN) can automatically extract discriminative features for fold recognition without human intervention, which has demonstrated an impressive performance on protein fold recognition. Despite the encouraging progress, DCNN often acts as a black box, and as such, it is challenging for users to understand what really happens in DCNN and why it works well for protein fold recognition. In this study, we explore the intrinsic mechanism of DCNN and explain why it works for protein fold recognition using a visual explanation technique. More specifically, we first trained a VGGNet-based DCNN model, termed VGGNet-FE, which can extract fold-specific features from the predicted protein residue-residue contact map for protein fold recognition. Subsequently, based on the trained VGGNet-FE, we implemented a new contact-assisted predictor, termed VGGfold, for protein fold recognition; we then visualized what features were extracted by each of the convolutional layers in VGGNet-FE using a deconvolution technique. Furthermore, we visualized the high-level semantic information, termed fold-discriminative region, of a predicted contact map from the localization map obtained from the last convolutional layer of VGGNet-FE. It is visually confirmed that VGGNet-FE could effectively extract distinct fold-discriminative regions for different types of protein folds, thereby accounting for the improved performance of VGGfold for protein fold recognition. In summary, this study is of great significance for both understanding the working principle of DCNNs in protein fold recognition and exploring the relationship between the predicted protein contact map and protein tertiary structure. This proposed visualization method is flexible and applicable to address other DCNN-based bioinformatics and computational biology questions. The online web server of VGGfold is freely available at http://csbio.njust.edu.cn/bioinf/vggfold/.

DeepDist: real-value inter-residue distance prediction with deep residual convolutional network

BACKGROUND

Driven by deep learning, inter-residue contact/distance prediction has been significantly improved and substantially enhanced ab initio protein structure prediction. Currently, most of the distance prediction methods classify inter-residue distances into multiple distance intervals instead of directly predicting real-value distances. The output of the former has to be converted into real-value distances to be used in tertiary structure prediction.

RESULTS

To explore the potentials of predicting real-value inter-residue distances, we develop a multi-task deep learning distance predictor (DeepDist) based on new residual convolutional network architectures to simultaneously predict real-value inter-residue distances and classify them into multiple distance intervals. Tested on 43 CASP13 hard domains, DeepDist achieves comparable performance in real-value distance prediction and multi-class distance prediction. The average mean square error (MSE) of DeepDist's real-value distance prediction is 0.896 Ų when filtering out the predicted distance ≥ 16 Å, which is lower than 1.003 Ų of DeepDist's multi-class distance prediction. When distance predictions are converted into contact predictions at 8 Å threshold (the standard threshold in the field), the precision of top L/5 and L/2 contact predictions of DeepDist's multi-class distance prediction is 79.3% and 66.1%, respectively, higher than 78.6% and 64.5% of its real-value distance prediction and the best results in the CASP13 experiment.

CONCLUSIONS

DeepDist can predict inter-residue distances well and improve binary contact prediction over the existing state-of-the-art methods. Moreover, the predicted real-value distances can be directly used to reconstruct protein tertiary structures better than multi-class distance predictions due to the lower MSE. Finally, we demonstrate that predicting the real-value distance map and multi-class distance map at the same time performs better than predicting real-value distances alone.

Probabilistic divergence of a template-based modelling methodology from the ideal protocol

Protein structural information is essential for the detailed mapping of a functional protein network. For a higher modelling accuracy and quicker implementation, template-based algorithms have been extensively deployed and redefined. The methods only assess the predicted structure against its native state/template and do not estimate the accuracy for each modelling step. A divergence measure is therefore postulated to estimate the modelling accuracy against its theoretical optimal benchmark. By freezing the domain boundaries, the divergence measures are predicted for the most crucial steps of a modelling algorithm. To precisely refine the score using weighting constants, big data analysis could further be deployed.

Crystallographic molecular replacement using an in silico-generated search model of SARS-CoV-2 ORF8

The majority of crystal structures are determined by the method of molecular replacement (MR). The range of application of MR is limited mainly by the need for an accurate search model. In most cases, pre-existing experimentally determined structures are used as search models. In favorable cases, ab initio predicted structures have yielded search models adequate for molecular replacement. The ORF8 protein of SARS-CoV-2 represents a challenging case for MR using an ab initio prediction because ORF8 has an all β-sheet fold and few orthologs. We previously determined experimentally the structure of ORF8 using the single anomalous dispersion (SAD) phasing method, having been unable to find an MR solution to the crystallographic phase problem. Following a report of an accurate prediction of the ORF8 structure, we assessed whether the predicted model would have succeeded as an MR search model. A phase problem solution was found, and the resulting structure was refined, yielding structural parameters equivalent to the original experimental solution.

A Review of Protein Structure Prediction using Deep Learning

Proteins are macromolecules composed of 20 types of amino acids in a specific order. Understanding how proteins fold is vital because its 3-dimensional structure determines the function of a protein. Prediction of protein structure based on amino acid strands and evolutionary information becomes the basis for other studies such as predicting the function, property or behaviour of a protein and modifying or designing new proteins to perform certain desired functions. Machine learning advances, particularly deep learning, are igniting a paradigm shift in scientific study. In this review, we summarize recent work in applying deep learning techniques to tackle problems in protein structural prediction. We discuss various deep learning approaches used to predict protein structure and future achievements and challenges. This review is expected to help provide perspectives on problems in biochemistry that can take advantage of the deep learning approach. Some of the unanswered challenges with current computational approaches are predicting the location and precision orientation of protein side chains, predicting protein interactions with DNA, RNA and other small molecules and predicting the structure of protein complexes.

Hybrid methods for combined experimental and computational determination of protein structure

Knowledge of protein structure is paramount to the understanding of biological function, developing new therapeutics, and making detailed mechanistic hypotheses. Therefore, methods to accurately elucidate three-dimensional structures of proteins are in high demand. While there are a few experimental techniques that can routinely provide high-resolution structures, such as x-ray crystallography, nuclear magnetic resonance (NMR), and cryo-EM, which have been developed to determine the structures of proteins, these techniques each have shortcomings and thus cannot be used in all cases. However, additionally, a large number of experimental techniques that provide some structural information, but not enough to assign atomic positions with high certainty have been developed. These methods offer sparse experimental data, which can also be noisy and inaccurate in some instances. In cases where it is not possible to determine the structure of a protein experimentally, computational structure prediction methods can be used as an alternative. Although computational methods can be performed without any experimental data in a large number of studies, inclusion of sparse experimental data into these prediction methods has yielded significant improvement. In this Perspective, we cover many of the successes of integrative modeling, computational modeling with experimental data, specifically for protein folding, protein-protein docking, and molecular dynamics simulations. We describe methods that incorporate sparse data from cryo-EM, NMR, mass spectrometry, electron paramagnetic resonance, small-angle x-ray scattering, Förster resonance energy transfer, and genetic sequence covariation. Finally, we highlight some of the major challenges in the field as well as possible future directions.

ITP-Pred: an interpretable method for predicting, therapeutic peptides with fused features low-dimension representation

The peptide therapeutics market is providing new opportunities for the biotechnology and pharmaceutical industries. Therefore, identifying therapeutic peptides and exploring their properties are important. Although several studies have proposed different machine learning methods to predict peptides as being therapeutic peptides, most do not explain the decision factors of model in detail. In this work, an Interpretable Therapeutic Peptide Prediction (ITP-Pred) model based on efficient feature fusion was developed. First, we proposed three kinds of feature descriptors based on sequence and physicochemical property encoded, namely amino acid composition (AAC), group AAC and coding autocorrelation, and concatenated them to obtain the feature representation of therapeutic peptide. Then, we input it into the CNN-Bi-directional Long Short-Term Memory (BiLSTM) model to automatically learn recognition of therapeutic peptides. The cross-validation and independent verification experiments results indicated that ITP-Pred has a higher prediction performance on the benchmark dataset than other comparison methods. Finally, we analyzed the output of the model from two aspects: sequence order and physical and chemical properties, mining important features as guidance for the design of better models that can complement existing methods.

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.

Homology modeling in the time of collective and artificial intelligence

Homology modeling is a method for building protein 3D structures using protein primary sequence and utilizing prior knowledge gained from structural similarities with other proteins. The homology modeling process is done in sequential steps where sequence/structure alignment is optimized, then a backbone is built and later, side-chains are added. Once the low-homology loops are modeled, the whole 3D structure is optimized and validated. In the past three decades, a few collective and collaborative initiatives allowed for continuous progress in both homology and ab initio modeling. Critical Assessment of protein Structure Prediction (CASP) is a worldwide community experiment that has historically recorded the progress in this field. Folding@Home and Rosetta@Home are examples of crowd-sourcing initiatives where the community is sharing computational resources, whereas RosettaCommons is an example of an initiative where a community is sharing a codebase for the development of computational algorithms. Foldit is another initiative where participants compete with each other in a protein folding video game to predict 3D structure. In the past few years, contact maps deep machine learning was introduced to the 3D structure prediction process, adding more information and increasing the accuracy of models significantly. In this review, we will take the reader in a journey of exploration from the beginnings to the most recent turnabouts, which have revolutionized the field of homology modeling. Moreover, we discuss the new trends emerging in this rapidly growing field.

Deep Learning in Proteomics

Proteomics, the study of all the proteins in biological systems, is becoming a data-rich science. Protein sequences and structures are comprehensively catalogued in online databases. With recent advancements in tandem mass spectrometry (MS) technology, protein expression and post-translational modifications (PTMs) can be studied in a variety of biological systems at the global scale. Sophisticated computational algorithms are needed to translate the vast amount of data into novel biological insights. Deep learning automatically extracts data representations at high levels of abstraction from data, and it thrives in data-rich scientific research domains. Here, a comprehensive overview of deep learning applications in proteomics, including retention time prediction, MS/MS spectrum prediction, de novo peptide sequencing, PTM prediction, major histocompatibility complex-peptide binding prediction, and protein structure prediction, is provided. Limitations and the future directions of deep learning in proteomics are also discussed. This review will provide readers an overview of deep learning and how it can be used to analyze proteomics data.

FilterDCA: Interpretable supervised contact prediction using inter-domain coevolution

Predicting three-dimensional protein structure and assembling protein complexes using sequence information belongs to the most prominent tasks in computational biology. Recently substantial progress has been obtained in the case of single proteins using a combination of unsupervised coevolutionary sequence analysis with structurally supervised deep learning. While reaching impressive accuracies in predicting residue-residue contacts, deep learning has a number of disadvantages. The need for large structural training sets limits the applicability to multi-protein complexes; and their deep architecture makes the interpretability of the convolutional neural networks intrinsically hard. Here we introduce FilterDCA, a simpler supervised predictor for inter-domain and inter-protein contacts. It is based on the fact that contact maps of proteins show typical contact patterns, which results from secondary structure and are reflected by patterns in coevolutionary analysis. We explicitly integrate averaged contacts patterns with coevolutionary scores derived by Direct Coupling Analysis, improving performance over standard coevolutionary analysis, while remaining fully transparent and interpretable. The FilterDCA code is available at http://gitlab.lcqb.upmc.fr/muscat/FilterDCA.

The de novo prediction of tertiary and quaternary protein structures has recently seen important advances, by combining unsupervised, purely sequence-based coevolutionary analyses with structure-based supervision using deep learning for contact-map prediction. While showing impressive performance, deep-learning methods require large training sets and pose severe obstacles for their interpretability. Here we construct a simple, transparent and therefore fully interpretable inter-domain contact predictor, which uses the results of coevolutionary Direct Coupling Analysis in combination with explicitly constructed filters reflecting typical contact patterns in a training set of known protein structures, and which improves the accuracy of predicted contacts significantly. Our approach thereby sheds light on the question how contact information is encoded in coevolutionary signals.

AWSEM-Suite: a protein structure prediction server based on template-guided, coevolutionary-enhanced optimized folding landscapes

The accurate and reliable prediction of the 3D structures of proteins and their assemblies remains difficult even though the number of solved structures soars and prediction techniques improve. In this study, a free and open access web server, AWSEM-Suite, whose goal is to predict monomeric protein tertiary structures from sequence is described. The model underlying the server’s predictions is a coarse-grained protein force field which has its roots in neural network ideas that has been optimized using energy landscape theory. Employing physically motivated potentials and knowledge-based local structure biasing terms, the addition of homologous template and co-evolutionary restraints to AWSEM-Suite greatly improves the predictive power of pure AWSEM structure prediction. From the independent evaluation metrics released in the CASP13 experiment, AWSEM-Suite proves to be a reasonably accurate algorithm for free modeling, standing at the eighth position in the free modeling category of CASP13. The AWSEM-Suite server also features a front end with a user-friendly interface. The AWSEM-Suite server is a powerful tool for predicting monomeric protein tertiary structures that is most useful when a suitable structure template is not available. The AWSEM-Suite server is freely available at: https://awsem.rice.edu.

DEEPCON: protein contact prediction using dilated convolutional neural networks with dropout

MOTIVATION

Exciting new opportunities have arisen to solve the protein contact prediction problem from the progress in neural networks and the availability of a large number of homologous sequences through high-throughput sequencing. In this work, we study how deep convolutional neural networks (ConvNets) may be best designed and developed to solve this long-standing problem.

RESULTS

With publicly available datasets, we designed and trained various ConvNet architectures. We tested several recent deep learning techniques including wide residual networks, dropouts and dilated convolutions. We studied the improvements in the precision of medium-range and long-range contacts, and compared the performance of our best architectures with the ones used in existing state-of-the-art methods. The proposed ConvNet architectures predict contacts with significantly more precision than the architectures used in several state-of-the-art methods. When trained using the DeepCov dataset consisting of 3456 proteins and tested on PSICOV dataset of 150 proteins, our architectures achieve up to 15% higher precision when L/2 long-range contacts are evaluated. Similarly, when trained using the DNCON2 dataset consisting of 1426 proteins and tested on 84 protein domains in the CASP12 dataset, our single network achieves 4.8% higher precision than the ensembled DNCON2 method when top L long-range contacts are evaluated.

AVAILABILITY AND IMPLEMENTATION

DEEPCON is available at https://github.com/badriadhikari/DEEPCON/.

DeepMSA: constructing deep multiple sequence alignment to improve contact prediction and fold-recognition for distant-homology proteins

MOTIVATION

The success of genome sequencing techniques has resulted in rapid explosion of protein sequences. Collections of multiple homologous sequences can provide critical information to the modeling of structure and function of unknown proteins. There are however no standard and efficient pipeline available for sensitive multiple sequence alignment (MSA) collection. This is particularly challenging when large whole-genome and metagenome databases are involved.

RESULTS

We developed DeepMSA, a new open-source method for sensitive MSA construction, which has homologous sequences and alignments created from multi-sources of whole-genome and metagenome databases through complementary hidden Markov model algorithms. The practical usefulness of the pipeline was examined in three large-scale benchmark experiments based on 614 non-redundant proteins. First, DeepMSA was utilized to generate MSAs for residue-level contact prediction by six coevolution and deep learning-based programs, which resulted in an accuracy increase in long-range contacts by up to 24.4% compared to the default programs. Next, multiple threading programs are performed for homologous structure identification, where the average TM-score of the template alignments has over 7.5% increases with the use of the new DeepMSA profiles. Finally, DeepMSA was used for secondary structure prediction and resulted in statistically significant improvements in the Q3 accuracy. It is noted that all these improvements were achieved without re-training the parameters and neural-network models, demonstrating the robustness and general usefulness of the DeepMSA in protein structural bioinformatics applications, especially for targets without homologous templates in the PDB library.

AVAILABILITY AND IMPLEMENTATION

https://zhanglab.ccmb.med.umich.edu/DeepMSA/.

SUPPLEMENTARY INFORMATION

Supplementary data are available at Bioinformatics online.

A novel fusion based on the evolutionary features for protein fold recognition using support vector machines

Protein fold recognition plays a crucial role in discovering three-dimensional structure of proteins and protein functions. Several approaches have been employed for the prediction of protein folds. Some of these approaches are based on extracting features from protein sequences and using a strong classifier. Feature extraction techniques generally utilize syntactical-based information, evolutionary-based information and physicochemical-based information to extract features. In recent years, finding an efficient technique for integrating discriminate features have been received advancing attention. In this study, we integrate Auto-Cross-Covariance and Separated dimer evolutionary feature extraction methods. The results’ features are scored by Information gain to define and select several discriminated features. According to three benchmark datasets, DD, RDD ,and EDD, the results of the support vector machine show more than 6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\%$$\end{document}% improvement in accuracy on these benchmark datasets.

A fully open-source framework for deep learning protein real-valued distances

As deep learning algorithms drive the progress in protein structure prediction, a lot remains to be studied at this merging superhighway of deep learning and protein structure prediction. Recent findings show that inter-residue distance prediction, a more granular version of the well-known contact prediction problem, is a key to predicting accurate models. However, deep learning methods that predict these distances are still in the early stages of their development. To advance these methods and develop other novel methods, a need exists for a small and representative dataset packaged for faster development and testing. In this work, we introduce protein distance net (PDNET), a framework that consists of one such representative dataset along with the scripts for training and testing deep learning methods. The framework also includes all the scripts that were used to curate the dataset, and generate the input features and distance maps. Deep learning models can also be trained and tested in a web browser using free platforms such as Google Colab. We discuss how PDNET can be used to predict contacts, distance intervals, and real-valued distances.

DeepHelicon: Accurate prediction of inter-helical residue contacts in transmembrane proteins by residual neural networks

Accurate prediction of amino acid residue contacts is an important prerequisite for generating high-quality 3D models of transmembrane (TM) proteins. While a large number of compositional, evolutionary, and structural properties of proteins can be used to train contact prediction methods, recent research suggests that coevolution between residues provides the strongest indication of their spatial proximity. We have developed a deep learning approach, DeepHelicon, to predict inter-helical residue contacts in TM proteins by considering only coevolutionary features. DeepHelicon comprises a two-stage supervised learning process by residual neural networks for a gradual refinement of contact maps, followed by variance reduction by an ensemble of models. We present a benchmark study of 12 contact predictors and conclude that DeepHelicon together with the two other state-of-the-art methods DeepMetaPSICOV and Membrain2 outperforms the 10 remaining algorithms on all datasets and at all settings. On a set of 44 TM proteins with an average length of 388 residues DeepHelicon achieves the best performance among all benchmarked methods in predicting the top L/5 and L/2 inter-helical contacts, with the mean precision of 87.42% and 77.84%, respectively. On a set of 57 relatively small TM proteins with an average length of 298 residues DeepHelicon ranks second best after DeepMetaPSICOV. DeepHelicon produces the most accurate predictions for large proteins with more than 10 transmembrane helices. Coevolutionary features alone allow to predict inter-helical residue contacts with an accuracy sufficient for generating acceptable 3D models for up to 30% of proteins using a fully automated modeling method such as CONFOLD2.

Artificial Intelligence Teaches Drugs to Target Proteins by Tackling the Induced Folding Problem

We explore the possibility of a deep learning (DL) platform that steers drug design to target proteins by inducing binding-competent conformations. We deal with the fact that target proteins are usually not fixed targets but structurally adapt to the ligand in ways that need to be predicted to enable pharmaceutical discovery. In contrast with protein folding predictors, the proposed DL system integrates signals for structural disorder to predict conformations in floppy regions of the target protein that rely on associations with the purposely designed drug to maintain their structural integrity. This is tantamount to solve the drug-induced folding problem within an AI-empowered drug discovery platform. Preliminary testing of the proposed DL platform reveals that it is possible to infer the induced folding ensemble from which a therapeutically targetable conformation gets selected by DL-instructed drug design.

ResPRE: high-accuracy protein contact prediction by coupling precision matrix with deep residual neural networks

MOTIVATION

Contact-map of a protein sequence dictates the global topology of structural fold. Accurate prediction of the contact-map is thus essential to protein 3D structure prediction, which is particularly useful for the protein sequences that do not have close homology templates in the Protein Data Bank.

RESULTS

We developed a new method, ResPRE, to predict residue-level protein contacts using inverse covariance matrix (or precision matrix) of multiple sequence alignments (MSAs) through deep residual convolutional neural network training. The approach was tested on a set of 158 non-homologous proteins collected from the CASP experiments and achieved an average accuracy of 50.6% in the top-L long-range contact prediction with L being the sequence length, which is 11.7% higher than the best of other state-of-the-art approaches ranging from coevolution coupling analysis to deep neural network training. Detailed data analyses show that the major advantage of ResPRE lies at the utilization of precision matrix that helps rule out transitional noises of contact-maps compared with the previously used covariance matrix. Meanwhile, the residual network with parallel shortcut layer connections increases the learning ability of deep neural network training. It was also found that appropriate collection of MSAs can further improve the accuracy of final contact-map predictions. The standalone package and online server of ResPRE are made freely available, which should bring important impact on protein structure and function modeling studies in particular for the distant- and non-homology protein targets.

AVAILABILITY AND IMPLEMENTATION

https://zhanglab.ccmb.med.umich.edu/ResPRE and https://github.com/leeyang/ResPRE.

SUPPLEMENTARY INFORMATION

Supplementary data are available at Bioinformatics online.

Protein homodimers structure prediction based on deep neural network

Structural prediction of protein-protein complexes has important application in such domains as modeling of biological processes and drug design. Homodimers (complexes which consist of two identical proteins) are the most common type of protein complexes in nature but there is still no universal algorithm to predict their 3D structures. Experimental techniques to identify the structure of protein complex require enormous amount of time and resources, and each method has its own limitations. Recently Deep Neural Networks allowed to predict structures of individual proteins greatly prevailing in accuracy over other algorithmic approaches. Building on the idea of this approach, we developed an algorithm to model the 3D structure of homodimer based on deep learning. It consists of two major steps: at the first step a protein complex contact map is predicted with the deep convolutional neural network, and the second stage is used to predict 3D structure of homodimer based on obtained contact map and optimization procedure. The use of the neural network in combination with optimization procedure based on gradient descent method allowed to predict structures for protein homodimers. The suggested approach was tested and validated on a dataset of protein homodimers from Protein Data Bank (PDB). The developed procedure could be also used for evaluating protein homodimer models as one of the stages in drug compounds developing.

GalaxyRefine2: simultaneous refinement of inaccurate local regions and overall protein structure

The 3D structure of a protein can be predicted from its amino acid sequence with high accuracy for a large fraction of cases because of the availability of large quantities of experimental data and the advance of computational algorithms. Recently, deep learning methods exploiting the coevolution information obtained by comparing related protein sequences have been successfully used to generate highly accurate model structures even in the absence of template structure information. However, structures predicted based on either template structures or related sequences require further improvement in regions for which information is missing. Refining a predicted protein structure with insufficient information on certain regions is critical because these regions may be connected to functional specificity that is not conserved among related proteins. The GalaxyRefine2 web server, freely available via http://galaxy.seoklab.org/refine2, is an upgraded version of the GalaxyRefine protein structure refinement server and reflects recent developments successfully tested through CASP blind prediction experiments. This method adopts an iterative optimization approach involving various structure move sets to refine both local and global structures. The estimation of local error and hybridization of available homolog structures are also employed for effective conformation search.

Protein Contact Map Prediction Based on ResNet and DenseNet

Residue-residue contact prediction has become an increasingly important tool for modeling the three-dimensional structure of a protein when no homologous structure is available. Ultradeep residual neural network (ResNet) has become the most popular method for making contact predictions because it captures the contextual information between residues. In this paper, we propose a novel deep neural network framework for contact prediction which combines ResNet and DenseNet. This framework uses 1D ResNet to process sequential features, and besides PSSM, SS3, and solvent accessibility, we have introduced a new feature, position-specific frequency matrix (PSFM), as an input. Using ResNet's residual module and identity mapping, it can effectively process sequential features after which the outer concatenation function is used for sequential and pairwise features. Prediction accuracy is improved following a final processing step using the dense connection of DenseNet. The prediction accuracy of the protein contact map shows that our method is more effective than other popular methods due to the new network architecture and the added feature input.

Artificial Intelligence Steering Molecular Therapy in the Absence of Information on Target Structure and Regulation

Protein associations are at the core of biological activity, and the drug-based disruption of dysfunctional associations poses a major challenge to targeted therapy. The problem becomes daunting when the structure and regulated modulation of the complex are unknown. To address the challenge, we leverage an artificial intelligence platform that learns from structural and epistructural data and infers regulation-susceptible regions that also generate interfacial tension between protein and water, thereby promoting protein associations. The input consists of sequence-derived 1D-features. The network is configured with evolutionarily coupled residues and taught to search for phosphorylation-modulated binding epitopes. The discovery platform is benchmarked against a PDB-derived testing set and validated against experimental data on a therapeutic disruptor designed according to the inferred epitope for a large deregulated complex known to be recruited in heart failure. Thus, dysfunctional "molecular brakes" of cardiac contractility get released through a therapeutic intervention guided by artificial intelligence.

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.

Improved protein structure prediction using potentials from deep learning

Protein structure prediction can be used to determine the three-dimensional shape of a protein from its amino acid sequence¹. This problem is of fundamental importance as the structure of a protein largely determines its function²; however, protein structures can be difficult to determine experimentally. Considerable progress has recently been made by leveraging genetic information. It is possible to infer which amino acid residues are in contact by analysing covariation in homologous sequences, which aids in the prediction of protein structures³. Here we show that we can train a neural network to make accurate predictions of the distances between pairs of residues, which convey more information about the structure than contact predictions. Using this information, we construct a potential of mean force⁴ that can accurately describe the shape of a protein. We find that the resulting potential can be optimized by a simple gradient descent algorithm to generate structures without complex sampling procedures. The resulting system, named AlphaFold, achieves high accuracy, even for sequences with fewer homologous sequences. In the recent Critical Assessment of Protein Structure Prediction⁵ (CASP13)-a blind assessment of the state of the field-AlphaFold created high-accuracy structures (with template modelling (TM) scores⁶ of 0.7 or higher) for 24 out of 43 free modelling domains, whereas the next best method, which used sampling and contact information, achieved such accuracy for only 14 out of 43 domains. AlphaFold represents a considerable advance in protein-structure prediction. We expect this increased accuracy to enable insights into the function and malfunction of proteins, especially in cases for which no structures for homologous proteins have been experimentally determined⁷.

DeepECA: an end-to-end learning framework for protein contact prediction from a multiple sequence alignment

Background

Recently developed methods of protein contact prediction, a crucially important step for protein structure prediction, depend heavily on deep neural networks (DNNs) and multiple sequence alignments (MSAs) of target proteins. Protein sequences are accumulating to an increasing degree such that abundant sequences to construct an MSA of a target protein are readily obtainable. Nevertheless, many cases present different ends of the number of sequences that can be included in an MSA used for contact prediction. The abundant sequences might degrade prediction results, but opportunities remain for a limited number of sequences to construct an MSA. To resolve these persistent issues, we strove to develop a novel framework using DNNs in an end-to-end manner for contact prediction.

Results

We developed neural network models to improve precision of both deep and shallow MSAs. Results show that higher prediction accuracy was achieved by assigning weights to sequences in a deep MSA. Moreover, for shallow MSAs, adding a few sequential features was useful to increase the prediction accuracy of long-range contacts in our model. Based on these models, we expanded our model to a multi-task model to achieve higher accuracy by incorporating predictions of secondary structures and solvent-accessible surface areas. Moreover, we demonstrated that ensemble averaging of our models can raise accuracy. Using past CASP target protein domains, we tested our models and demonstrated that our final model is superior to or equivalent to existing meta-predictors.

Conclusions

The end-to-end learning framework we built can use information derived from either deep or shallow MSAs for contact prediction. Recently, an increasing number of protein sequences have become accessible, including metagenomic sequences, which might degrade contact prediction results. Under such circumstances, our model can provide a means to reduce noise automatically. According to results of tertiary structure prediction based on contacts and secondary structures predicted by our model, more accurate three-dimensional models of a target protein are obtainable than those from existing ECA methods, starting from its MSA. DeepECA is available from https://github.com/tomiilab/DeepECA.

Deep learning for mining protein data

The recent emergence of deep learning to characterize complex patterns of protein big data reveals its potential to address the classic challenges in the field of protein data mining. Much research has revealed the promise of deep learning as a powerful tool to transform protein big data into valuable knowledge, leading to scientific discoveries and practical solutions. In this review, we summarize recent publications on deep learning predictive approaches in the field of mining protein data. The application architectures of these methods include multilayer perceptrons, stacked autoencoders, deep belief networks, two- or three-dimensional convolutional neural networks, recurrent neural networks, graph neural networks, and complex neural networks and are described from five perspectives: residue-level prediction, sequence-level prediction, three-dimensional structural analysis, interaction prediction, and mass spectrometry data mining. The advantages and deficiencies of these architectures are presented in relation to various tasks in protein data mining. Additionally, some practical issues and their future directions are discussed, such as robust deep learning for protein noisy data, architecture optimization for specific tasks, efficient deep learning for limited protein data, multimodal deep learning for heterogeneous protein data, and interpretable deep learning for protein understanding. This review provides comprehensive perspectives on general deep learning techniques for protein data analysis.

Deep-learning contact-map guided protein structure prediction in CASP13

We report the results of two fully automated structure prediction pipelines, "Zhang-Server" and "QUARK", in CASP13. The pipelines were built upon the C-I-TASSER and C-QUARK programs, which in turn are based on I-TASSER and QUARK but with three new modules: (a) a novel multiple sequence alignment (MSA) generation protocol to construct deep sequence-profiles for contact prediction; (b) an improved meta-method, NeBcon, which combines multiple contact predictors, including ResPRE that predicts contact-maps by coupling precision-matrices with deep residual convolutional neural-networks; and (c) an optimized contact potential to guide structure assembly simulations. For 50 CASP13 FM domains that lacked homologous templates, average TM-scores of the first models produced by C-I-TASSER and C-QUARK were 28% and 56% higher than those constructed by I-TASSER and QUARK, respectively. For the first time, contact-map predictions demonstrated usefulness on TBM domains with close homologous templates, where TM-scores of C-I-TASSER models were significantly higher than those of I-TASSER models with a P-value <.05. Detailed data analyses showed that the success of C-I-TASSER and C-QUARK was mainly due to the increased accuracy of deep-learning-based contact-maps, as well as the careful balance between sequence-based contact restraints, threading templates, and generic knowledge-based potentials. Nevertheless, challenges still remain for predicting quaternary structure of multi-domain proteins, due to the difficulties in domain partitioning and domain reassembly. In addition, contact prediction in terminal regions was often unsatisfactory due to the sparsity of MSAs. Development of new contact-based domain partitioning and assembly methods and training contact models on sparse MSAs may help address these issues.

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.