Protein tertiary structure modeling driven by deep learning and contact distance prediction in CASP13

Predicting residue-residue distance relationships (eg, contacts) has become the key direction to advance protein structure prediction since 2014 CASP11 experiment, while deep learning has revolutionized the technology for contact and distance distribution prediction since its debut in 2012 CASP10 experiment. During 2018 CASP13 experiment, we enhanced our MULTICOM protein structure prediction system with three major components: contact distance prediction based on deep convolutional neural networks, distance-driven template-free (ab initio) modeling, and protein model ranking empowered by deep learning and contact prediction. Our experiment demonstrates that contact distance prediction and deep learning methods are the key reasons that MULTICOM was ranked 3rd out of all 98 predictors in both template-free and template-based structure modeling in CASP13. Deep convolutional neural network can utilize global information in pairwise residue-residue features such as coevolution scores to substantially improve contact distance prediction, which played a decisive role in correctly folding some free modeling and hard template-based modeling targets. Deep learning also successfully integrated one-dimensional structural features, two-dimensional contact information, and three-dimensional structural quality scores to improve protein model quality assessment, where the contact prediction was demonstrated to consistently enhance ranking of protein models for the first time. The success of MULTICOM system clearly shows that protein contact distance prediction and model selection driven by deep learning holds the key of solving protein structure prediction problem. However, there are still challenges in accurately predicting protein contact distance when there are few homologous sequences, folding proteins from noisy contact distances, and ranking models of hard targets.     

Thiol-exchange in DTSSP crosslinked peptides is proportional to cysteine content and precisely controlled in crosslink detection by two-step LC-MALDI MSMS

The lysine-specific crosslinker 3,3'-dithiobis(sulfosuccinimidylpropionate) (DTSSP) is commonly used in the structural characterization of proteins by chemical crosslinking and mass spectrometry and we here describe an efficient two-step LC-MALDI-TOF/TOF procedure to detect crosslinked peptides. First MS data are acquired, and the properties of isotope-labeled DTSSP are used in data analysis to identify candidate crosslinks. MSMS data are then acquired for a restricted number of precursor ions per spot for final crosslink identification. We show that the thiol-catalyzed exchange between crosslinked peptides, which is due to the disulfide bond in DTSSP and known to possibly obscure data, can be precisely quantified using isotope-labeled DTSSP. Crosslinked peptides are recognized as 8 Da doublet peaks and a new isotopic peak with twice the intensity appears in the middle of the doublet as a consequence of the thiol-exchange. False-positive crosslinks, formed exclusively by thiol-exchange, yield a 1:2:1 isotope pattern, whereas true crosslinks, formed by two lysine residues within crosslinkable distance in the native protein structure, yield a 1:0:1 isotope pattern. Peaks with a 1:X:1 isotope pattern, where 0 < X < 2, can be trusted as true crosslinks, with a defined proportion of the signal [2X/(2 + X)] being noise from the thiol-exchange. The thiol-exchange was correlated with the protein cysteine content and was minimized by shortening the trypsin incubation time, and for two molecular chaperone proteins with known structure all crosslinks fitted well to the structure data. The thiol-exchange can thus be controlled and isotope-labeled DTSSP safely used to detect true crosslinks between lysine residues in proteins.     

Improving protein tertiary structure prediction by deep learning and distance prediction in CASP14

Substantial progresses in protein structure prediction have been made by utilizing deep-learning and residue-residue distance prediction since CASP13. Inspired by the advances, we improve our CASP14 MULTICOM protein structure prediction system by incorporating three new components: (a) a new deep learning-based protein inter-residue distance predictor to improve template-free (ab initio) tertiary structure prediction, (b) an enhanced template-based tertiary structure prediction method, and (c) distance-based model quality assessment methods empowered by deep learning. In the 2020 CASP14 experiment, MULTICOM predictor was ranked seventh out of 146 predictors in tertiary structure prediction and ranked third out of 136 predictors in inter-domain structure prediction. The results demonstrate that the template-free modeling based on deep learning and residue-residue distance prediction can predict the correct topology for almost all template-based modeling targets and a majority of hard targets (template-free targets or targets whose templates cannot be recognized), which is a significant improvement over the CASP13 MULTICOM predictor. Moreover, the template-free modeling performs better than the template-based modeling on not only hard targets but also the targets that have homologous templates. The performance of the template-free modeling largely depends on the accuracy of distance prediction closely related to the quality of multiple sequence alignments. The structural model quality assessment works well on targets for which enough good models can be predicted, but it may perform poorly when only a few good models are predicted for a hard target and the distribution of model quality scores is highly skewed. MULTICOM is available at https://github.com/jianlin-cheng/MULTICOM_Human_CASP14/tree/CASP14_DeepRank3 and https://github.com/multicom-toolbox/multicom/tree/multicom_v2.0 .     

Assessment of chemical-crosslink-assisted protein structure modeling in CASP13

With the advance of experimental procedures obtaining chemical crosslinking information is becoming a fast and routine practice. Information on crosslinks can greatly enhance the accuracy of protein structure modeling. Here, we review the current state of the art in modeling protein structures with the assistance of experimentally determined chemical crosslinks within the framework of the 13th meeting of Critical Assessment of Structure Prediction approaches. This largest-to-date blind assessment reveals benefits of using data assistance in difficult to model protein structure prediction cases. However, in a broader context, it also suggests that with the unprecedented advance in accuracy to predict contacts in recent years, experimental crosslinks will be useful only if their specificity and accuracy further improved and they are better integrated into computational workflows.     

Distance restraints from crosslinking mass spectrometry: Mining a molecular dynamics simulation database to evaluate lysine–lysine distances

Integrative structural biology attempts to model the structures of protein complexes that are challenging or intractable by classical structural methods (due to size, dynamics, or heterogeneity) by combining computational structural modeling with data from experimental methods. One such experimental method is chemical crosslinking mass spectrometry (XL‐MS), in which protein complexes are crosslinked and characterized using liquid chromatography‐mass spectrometry to pinpoint specific amino acid residues in close structural proximity. The commonly used lysine‐reactive N‐hydroxysuccinimide ester reagents disuccinimidylsuberate (DSS) and bis(sulfosuccinimidyl)suberate (BS³) have a linker arm that is 11.4 Å long when fully extended, allowing C_α (alpha carbon of protein backbone) atoms of crosslinked lysine residues to be up to ～24 Å apart. However, XL‐MS studies on proteins of known structure frequently report crosslinks that exceed this distance. Typically, a tolerance of ～3 Å is added to the theoretical maximum to account for this observation, with limited justification for the chosen value. We used the Dynameomics database, a repository of high‐quality molecular dynamics simulations of 807 proteins representative of diverse protein folds, to investigate the relationship between lysine–lysine distances in experimental starting structures and in simulation ensembles. We conclude that for DSS/BS³, a distance constraint of 26–30 Å between C_α atoms is appropriate. This analysis provides a theoretical basis for the widespread practice of adding a tolerance to the crosslinker length when comparing XL‐MS results to structures or in modeling. We also discuss the comparison of XL‐MS results to MD simulations and known structures as a means to test and validate experimental XL‐MS methods.     

Gene ontology improves template selection in comparative protein docking

Structural characterization of protein-protein interactions is essential for our ability to study life processes at the molecular level. Computational modeling of protein complexes (protein docking) is important as the source of their structure and as a way to understand the principles of protein interaction. Rapidly evolving comparative docking approaches utilize target/template similarity metrics, which are often based on the protein structure. Although the structural similarity, generally, yields good performance, other characteristics of the interacting proteins (eg, function, biological process, and localization) may improve the prediction quality, especially in the case of weak target/template structural similarity. For the ranking of a pool of models for each target, we tested scoring functions that quantify similarity of Gene Ontology (GO) terms assigned to target and template proteins in three ontology domains—biological process, molecular function, and cellular component (GO-score). The scoring functions were tested in docking of bound, unbound, and modeled proteins. The results indicate that the combined structural and GO-terms functions improve the scoring, especially in the twilight zone of structural similarity, typical for protein models of limited accuracy.     

Assessment of protein model structure accuracy estimation in CASP13: Challenges in the era of deep learning

Scoring model structure is an essential component of protein structure prediction that can affect the prediction accuracy tremendously. Users of protein structure prediction results also need to score models to select the best models for their application studies. In Critical Assessment of techniques for protein Structure Prediction (CASP), model accuracy estimation methods have been tested in a blind fashion by providing models submitted by the tertiary structure prediction servers for scoring. In CASP13, model accuracy estimation results were evaluated in terms of both global and local structure accuracy. Global structure accuracy estimation was evaluated by the quality of the models selected by the global structure scores and by the absolute estimates of the global scores. Residue-wise, local structure accuracy estimations were evaluated by three different measures. A new measure introduced in CASP13 evaluates the ability to predict inaccurately modeled regions that may be improved by refinement. An intensive comparative analysis on CASP13 and the previous CASPs revealed that the tertiary structure models generated by the CASP13 servers show very distinct features. Higher consensus toward models of higher global accuracy appeared even for free modeling targets, and many models of high global accuracy were not well optimized at the atomic level. This is related to the new technology in CASP13, deep learning for tertiary contact prediction. The tertiary model structures generated by deep learning pose a new challenge for EMA (estimation of model accuracy) method developers. Model accuracy estimation itself is also an area where deep learning can potentially have an impact, although current EMA methods have not fully explored that direction.     

Deep learning methods for 3D structural proteome and interactome modeling

Bolstered by recent methodological and hardware advances, deep learning has increasingly been applied to biological problems and structural proteomics. Such approaches have achieved remarkable improvements over traditional machine learning methods in tasks ranging from protein contact map prediction to protein folding, prediction of protein–protein interaction interfaces, and characterization of protein–drug binding pockets. In particular, emergence of ab initio protein structure prediction methods including AlphaFold2 has revolutionized protein structural modeling. From a protein function perspective, numerous deep learning methods have facilitated deconvolution of the exact amino acid residues and protein surface regions responsible for binding other proteins or small molecule drugs. In this review, we provide a comprehensive overview of recent deep learning methods applied in structural proteomics.     

Crosslinking of phycobiliproteins from the cyanobacterium Mastigocladus laminosus with bis-imidates: localization of an intrasubunit and an intersubunit crosslink in C-phycocyanin

The light-harvesting pigment-protein complexes allophycocyanin (AP), C-phycocyanin (PC) and phycoerythrocyanin (PEC) of the cyanobacterium Mastigocladus laminosus consist of alpha- and beta-subunits containing about 170 amino-acid residues each. These two subunits form an alpha,beta-monomer, three of which build up a disc-shaped trimer. In this study these phycobiliproteins were crosslinked with bis-imidates. Various spacer lengths of the reagent and various aggregation states of the phycobiliprotein were tested. An intersubunit crosslink could be verified in all three phycobiliproteins. PC-trimers were crosslinked with the homobifunctional reagent dimethyl pimelimidate having a maximal crosslinking distance of 10 A. Two crosslinks could be identified: an intramonomer intersubunit crosslink with a yield of 48% and an intrasubunit crosslink within alpha PC (57%). These products were chemically and enzymatically fragmented and the small crosslinked peptides were isolated and then identified by amino-acid analysis. The following amino acids were crosslinked: alpha-Val 1 with beta-Ala 1 and alpha-Lys 62 with alpha-Lys 134. Both crosslinks could be localized within the known three-dimensional structure of PC.     

Predicting interacting and interfacial residues using continuous sequence segments

Development of sequence-based methods for predicting putative interfacial residues is an extremely important task in modeling 3D structures of protein–protein complexes. In the present paper we used non-gapped sequence segments to predict both interacting and interfacial residues. We demonstrated that continuous sequence segments do occur at the protein–protein interfaces and showed that continuous interacting interfacial segments (CIIS) of length nine are presented on average, in 37% of the complexes in our dataset. Our results indicate that CIIS consist mostly of interacting strands and/or loops, while the CIIS involving the helixes are scarce. We performed scoring of CIIS using four different scoring mechanisms and found that scores of CIIS differ significantly from the scores calculated for random stretches of residues. We argue that such statistical difference inferred thought the corresponding Z-scores could be used for detecting putative interfacial residue segments without using any structural information. This hypothesis was tested on our dataset and benchmarking resulted to 10–60% prediction accuracy depending on type of benchmarking and scoring scheme used in calculations. Such predictions that do not depend on the availability of the 3D structures of monomers can be quite valuable in modeling 3D structures of obligatory complexes, for which structures of separated monomers do not exist.     

Predicting protein complex geometries with a neural network

A major challenge of the protein docking problem is to define scoring functions that can distinguish near‐native protein complex geometries from a large number of non‐native geometries (decoys) generated with noncomplexed protein structures (unbound docking). In this study, we have constructed a neural network that employs the information from atom‐pair distance distributions of a large number of decoys to predict protein complex geometries. We found that docking prediction can be significantly improved using two different types of polar hydrogen atoms. To train the neural network, 2000 near‐native decoys of even distance distribution were used for each of the 185 considered protein complexes. The neural network normalizes the information from different protein complexes using an additional protein complex identity input neuron for each complex. The parameters of the neural network were determined such that they mimic a scoring funnel in the neighborhood of the native complex structure. The neural network approach avoids the reference state problem, which occurs in deriving knowledge‐based energy functions for scoring. We show that a distance‐dependent atom pair potential performs much better than a simple atom‐pair contact potential. We have compared the performance of our scoring function with other empirical and knowledge‐based scoring functions such as ZDOCK 3.0, ZRANK, ITScore‐PP, EMPIRE, and RosettaDock. In spite of the simplicity of the method and its functional form, our neural network‐based scoring function achieves a reasonable performance in rigid‐body unbound docking of proteins. Proteins 2010. © 2009 Wiley‐Liss, Inc.     

基于深度学习的蛋白质建模与设计

随着蛋白质序列及结构数据的大量累积,在获得了大量描述性信息之后如何有效利用海量数据,从已有数据中高效提取信息并且应用到下游任务当中就成为了研究者亟待解决的问题。蛋白质的设计可使新蛋白的研发不再受限于实验条件,这对药物靶点预测、新药研发和材料设计等领域具有重要意义。深度学习作为一种高效的数据特征提取方法,可以通过它对蛋白质数据进行建模,进而加入先验信息对蛋白质进行设计。故此基于深度学习的蛋白质设计就成为一个具有广阔前景的研究领域。文中主要阐述基于深度学习的蛋白质序列与结构数据的建模和设计方法。详述该方法的策略、原理、适用范围、应用实例。讨论了深度学习方法在本领域的应用前景及局限性,以期为相关研究提供参考。     

Structural motifs in protein cores and at protein–protein interfaces are different

Structures of proteins and protein–protein complexes are determined by the same physical principles and thus share a number of similarities. At the same time, there could be differences because in order to function, proteins interact with other molecules, undergo conformations changes, and so forth, which might impose different restraints on the tertiary versus quaternary structures. This study focuses on structural properties of protein–protein interfaces in comparison with the protein core, based on the wealth of currently available structural data and new structure‐based approaches. The results showed that physicochemical characteristics, such as amino acid composition, residue–residue contact preferences, and hydrophilicity/hydrophobicity distributions, are similar in protein core and protein–protein interfaces. On the other hand, characteristics that reflect the evolutionary pressure, such as structural composition and packing, are largely different. The results provide important insight into fundamental properties of protein structure and function. At the same time, the results contribute to better understanding of the ways to dock proteins. Recent progress in predicting structures of individual proteins follows the advancement of deep learning techniques and new approaches to residue coevolution data. Protein core could potentially provide large amounts of data for application of the deep learning to docking. However, our results showed that the core motifs are significantly different from those at protein–protein interfaces, and thus may not be directly useful for docking. At the same time, such difference may help to overcome a major obstacle in application of the coevolutionary data to docking—discrimination of the intramolecular information not directly relevant to docking.     

Human protein aging: modification and crosslinking through dehydroalanine and dehydrobutyrine intermediates

Nonenzymatic post‐translational modification (PTM) of proteins is a fundamental molecular process of aging. The combination of various modifications and their accumulation with age not only affects function, but leads to crosslinking and protein aggregation. In this study, aged human lens proteins were examined using HPLC–tandem mass spectrometry and a blind PTM search strategy. Multiple thioether modifications of Ser and Thr residues by glutathione (GSH) and its metabolites were unambiguously identified. Thirty‐four of 36 sites identified on 15 proteins were found on known phosphorylation sites, supporting a mechanism involving dehydroalanine (DHA) and dehydrobutyrine (DHB) formation through β‐elimination of phosphoric acid from phosphoserine and phosphothreonine with subsequent nucleophilic attack by GSH. In vitro incubations of phosphopeptides demonstrated that this process can occur spontaneously under physiological conditions. Evidence that this mechanism can also lead to protein–protein crosslinks within cells is provided where five crosslinked peptides were detected in a human cataractous lens. Nondisulfide crosslinks were identified for the first time in lens tissue between βB2‐ & βB2‐, βA4‐ & βA3‐, γS‐ & βB1‐, and βA4‐ & βA4‐crystallins and provide detailed structural information on in vivo crystallin complexes. These data suggest that phosphoserine and phosphothreonine residues represent susceptible sites for spontaneous breakdown in long‐lived proteins and that DHA‐ and DHB‐mediated protein crosslinking may be the source of the long‐sought after nondisulfide protein aggregates believed to scatter light in cataractous lenses. Furthermore, this mechanism may be a common aging process that occurs in long‐lived proteins of other tissues leading to protein aggregation diseases.     

ScanNet: A Web Server for Structure-based Prediction of Protein Binding Sites with Geometric Deep Learning

Predicting the various binding sites of a protein from its structure sheds light on its function and paves the way towards design of interaction inhibitors. Here, we report ScanNet, a freely available web server for prediction of protein–protein, protein - disordered protein and protein - antibody binding sites from structure. ScanNet (Spatio-Chemical Arrangement of Neighbors Network) is an end-to-end, interpretable geometric deep learning model that learns spatio-chemical patterns directly from 3D structures. ScanNet consistently outperforms Machine Learning models based on handcrafted features and comparative modeling approaches. The web server is linked to both the PDB and AlphaFoldDB, and supports user-provided structure files. Predictions can be readily visualized on the website via the Molstar web app and locally via ChimeraX. ScanNet is available at http://bioinfo3d.cs.tau.ac.il/ScanNet/.     

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.     

Macromolecular docking restrained by a small angle X-ray scattering profile

While many structures of single protein components are becoming available, structural characterization of their complexes remains challenging. Methods for modeling assembly structures from individual components frequently suffer from large errors, due to protein flexibility and inaccurate scoring functions. However, when additional information is available, it may be possible to reduce the errors and compute near-native complex structures. One such type of information is a small angle X-ray scattering (SAXS) profile that can be collected in a high-throughput fashion from a small amount of sample in solution. Here, we present an efficient method for protein–protein docking with a SAXS profile (FoXSDock): generation of complex models by rigid global docking with PatchDock, filtering of the models based on the SAXS profile, clustering of the models, and refining the interface by flexible docking with FireDock. FoXSDock is benchmarked on 124 protein complexes with simulated SAXS profiles, as well as on 6 complexes with experimentally determined SAXS profiles. When induced fit is less than 1.5 Å interface C_α RMSD and the fraction residues of missing from the component structures is less than 3%, FoXSDock can find a model close to the native structure within the top 10 predictions in 77% of the cases; in comparison, docking alone succeeds in only 34% of the cases. Thus, the integrative approach significantly improves on molecular docking alone. The improvement arises from an increased resolution of rigid docking sampling and more accurate scoring.     

Structure prediction of protein complexes by an NMR-based protein docking algorithm

Protein docking algorithms can be used to study the driving forces and reaction mechanisms of docking processes. They are also able to speed up the lengthy process of experimental structure elucidation of protein complexes by proposing potential structures. In this paper, we are discussing a variant of the protein-protein docking problem, where the input consists of the tertiary structures of proteins A and B plus an unassigned one-dimensional ¹H-NMR spectrum of the complex AB. We present a new scoring function for evaluating and ranking potential complex structures produced by a docking algorithm. The scoring function computes a `theoretical' ¹H-NMR spectrum for each tentative complex structure and subtracts the calculated spectrum from the experimental one. The absolute areas of the difference spectra are then used to rank the potential complex structures. In contrast to formerly published approaches (e.g. [Morelli et al. (2000) Biochemistry, 39, 2530–2537]) we do not use distance constraints (intermolecular NOE constraints). We have tested the approach with four protein complexes whose three-dimensional structures are stored in the PDB data bank [Bernstein et al. (1977)] and whose ¹H-NMR shift assignments are available from the BMRB database. The best result was obtained for an example, where all standard scoring functions failed completely. Here, our new scoring function achieved an almost perfect separation between good approximations of the true complex structure and false positives.     

Folding of beta/alpha-unit scrambled forms of S. cerevisiae triosephosphate isomerase: Evidence for autonomy of substructure formation and plasticity of hydrophobic and hydrogen bonding interactions in core of (beta/alpha)8-barrel

We show that residues at the interfaces of protein-protein complexes have higher side-chain energy than other surface residues. Eight different sets of protein complexes were analyzed. For each protein pair, the complex structure was used to identify the interface residues in the unbound monomer structures. Side-chain energy was calculated for each surface residue in the unbound monomer using our previously developed scoring function.1 The mean energy was calculated for the interface residues and the other surface residues. In 15 of the 16 monomers, the mean energy of the interface residues was higher than that of other surface residues. By decomposing the scoring function, we found that the energy term of the buried surface area of non-hydrogen-bonded hydrophilic atoms is the most important factor contributing to the high energy of the interface regions. In spite of lacking hydrophilic residues, the interface regions were found to be rich in buried non-hydrogen-bonded hydrophilic atoms. Although the calculation results could be affected by the inaccuracy of the scoring function, patch analysis of side-chain energy on the surface of an isolated protein may be helpful in identifying the possible protein-protein interface. A patch was defined as 20 residues surrounding the central residue on the protein surface, and patch energy was calculated as the mean value of the side-chain energy of all residues in the patch. In 12 of the studied monomers, the patch with the highest energy overlaps with the observed interface. The results are more remarkable when only three residues with the highest energy in a patch are averaged to derive the patch energy. All three highest-energy residues of the top energy patch belong to interfacial residues in four of the eight small protomers. We also found that the residue with the highest energy score on the surface of a small protomer is very possibly the key interaction residue.     

Energy-based graph convolutional networks for scoring protein docking models

Structural information about protein-protein interactions, often missing at the interactome scale, is important for mechanistic understanding of cells and rational discovery of therapeutics. Protein docking provides a computational alternative for such information. However, ranking near-native docked models high among a large number of candidates, often known as the scoring problem, remains a critical challenge. Moreover, estimating model quality, also known as the quality assessment problem, is rarely addressed in protein docking. In this study, the two challenging problems in protein docking are regarded as relative and absolute scoring, respectively, and addressed in one physics-inspired deep learning framework. We represent protein and complex structures as intra- and inter-molecular residue contact graphs with atom-resolution node and edge features. And we propose a novel graph convolutional kernel that aggregates interacting nodes’ features through edges so that generalized interaction energies can be learned directly from 3D data. The resulting energy-based graph convolutional networks (EGCN) with multihead attention are trained to predict intra- and inter-molecular energies, binding affinities, and quality measures (interface RMSD) for encounter complexes. Compared to a state-of-the-art scoring function for model ranking, EGCN significantly improves ranking for a critical assessment of predicted interactions (CAPRI) test set involving homology docking; and is comparable or slightly better for Score_set, a CAPRI benchmark set generated by diverse community-wide docking protocols not known to training data. For Score_set quality assessment, EGCN shows about 27% improvement to our previous efforts. Directly learning from 3D structure data in graph representation, EGCN represents the first successful development of graph convolutional networks for protein docking.