Near-native Models Research Articles

DeepRank-GNN-esm: a graph neural network for scoring protein-protein models using protein language model.

Protein-Protein interactions (PPIs) play critical roles in numerous cellular processes. By modelling the 3D structures of the correspond protein complexes valuable insights can be obtained, providing, e.g. starting points for drug and protein design. One challenge in the modelling process is however the identification of near-native models from the large pool of generated models. To this end we have previously developed DeepRank-GNN, a graph neural network that integrates structural and sequence information to enable effective pattern learning at PPI interfaces. Its main features are related to the Position Specific Scoring Matrices (PSSMs), which are computationally expensive to generate, significantly limits the algorithm's usability. We introduce here DeepRank-GNN-esm that includes as additional features protein language model embeddings from the ESM-2 model. We show that the ESM-2 embeddings can actually replace the PSSM features at no cost in-, or even better performance on two PPI-related tasks: scoring docking poses and detecting crystal artifacts. This new DeepRank version bypasses thus the need of generating PSSM, greatly improving the usability of the software and opening new application opportunities for systems for which PSSM profiles cannot be obtained or are irrelevant (e.g. antibody-antigen complexes). DeepRank-GNN-esm is freely available from https://github.com/DeepRank/DeepRank-GNN-esm.

A Point Cloud-Based Deep Learning Model for Protein Docking Decoys Evaluation

Protein-protein docking reveals the process and product in protein interactions. Typically, a protein docking works with a docking model sampling, and then an evaluation method is used to rank the near-native models out from a large pool of generated decoys. In practice, the evaluation stage is the bottleneck to perform accurate protein docking. In this paper, PointNet, a deep learning algorithm based on point cloud, is applied to evaluate protein docking models. The proposed architecture is able to directly learn deep representations carrying the geometrical properties and atomic attributes from the 3D structural data of protein decoys. The experimental results show that the informative representations can benefit our proposed method to outperform other algorithms.

MetaScore: A Novel Machine-Learning-Based Approach to Improve Traditional Scoring Functions for Scoring Protein-Protein Docking Conformations.

Protein-protein interactions play a ubiquitous role in biological function. Knowledge of the three-dimensional (3D) structures of the complexes they form is essential for understanding the structural basis of those interactions and how they orchestrate key cellular processes. Computational docking has become an indispensable alternative to the expensive and time-consuming experimental approaches for determining the 3D structures of protein complexes. Despite recent progress, identifying near-native models from a large set of conformations sampled by docking-the so-called scoring problem-still has considerable room for improvement. We present MetaScore, a new machine-learning-based approach to improve the scoring of docked conformations. MetaScore utilizes a random forest (RF) classifier trained to distinguish near-native from non-native conformations using their protein-protein interfacial features. The features include physicochemical properties, energy terms, interaction-propensity-based features, geometric properties, interface topology features, evolutionary conservation, and also scores produced by traditional scoring functions (SFs). MetaScore scores docked conformations by simply averaging the score produced by the RF classifier with that produced by any traditional SF. We demonstrate that (i) MetaScore consistently outperforms each of the nine traditional SFs included in this work in terms of success rate and hit rate evaluated over conformations ranked among the top 10; (ii) an ensemble method, MetaScore-Ensemble, that combines 10 variants of MetaScore obtained by combining the RF score with each of the traditional SFs outperforms each of the MetaScore variants. We conclude that the performance of traditional SFs can be improved upon by using machine learning to judiciously leverage protein-protein interfacial features and by using ensemble methods to combine multiple scoring functions.

G-RANK: an equivariant graph neural network for the scoring of protein-protein docking models.

Protein complex structure prediction is important for many applications in bioengineering. A widely used method for predicting the structure of protein complexes is computational docking. Although many tools for scoring protein-protein docking models have been developed, it is still a challenge to accurately identify near-native models for unknown protein complexes. A recently proposed model called the geometric vector perceptron-graph neural network (GVP-GNN), a subtype of equivariant graph neural networks, has demonstrated success in various 3D molecular structure modeling tasks. Herein, we present G-RANK, a GVP-GNN-based method for the scoring of protein-protein docking models. When evaluated on two different test datasets, G-RANK achieved a performance competitive with or better than the state-of-the-art scoring functions. We expect G-RANK to be a useful tool for various applications in biological engineering. Source code is available at https://github.com/ha01994/grank. kds@kaist.ac.kr. Supplementary data are available at Bioinformatics Advances online.

Benchmarking AlphaFold for protein complex modeling reveals accuracy determinants.

High‐resolution experimental structural determination of protein–protein interactions has led to valuable mechanistic insights, yet due to the massive number of interactions and experimental limitations there is a need for computational methods that can accurately model their structures. Here we explore the use of the recently developed deep learning method, AlphaFold, to predict structures of protein complexes from sequence. With a benchmark of 152 diverse heterodimeric protein complexes, multiple implementations and parameters of AlphaFold were tested for accuracy. Remarkably, many cases (43%) had near‐native models (medium or high critical assessment of predicted interactions accuracy) generated as top‐ranked predictions by AlphaFold, greatly surpassing the performance of unbound protein–protein docking (9% success rate for near‐native top‐ranked models), however AlphaFold modeling of antibody–antigen complexes within our set was unsuccessful. We identified sequence and structural features associated with lack of AlphaFold success, and we also investigated the impact of multiple sequence alignment input. Benchmarking of a multimer‐optimized version of AlphaFold (AlphaFold‐Multimer) with a set of recently released antibody–antigen structures confirmed a low rate of success for antibody–antigen complexes (11% success), and we found that T cell receptor–antigen complexes are likewise not accurately modeled by that algorithm, showing that adaptive immune recognition poses a challenge for the current AlphaFold algorithm and model. Overall, our study demonstrates that end‐to‐end deep learning can accurately model many transient protein complexes, and highlights areas of improvement for future developments to reliably model any protein–protein interaction of interest.

Using steered molecular dynamic tension for assessing quality of computational protein structure models.

The native structures of proteins, except for notable exceptions of intrinsically disordered proteins, in general take their most stable conformation in the physiological condition to maintain their structural framework so that their biological function can be properly carried out. Experimentally, the stability of a protein can be measured by several means, among which the pulling experiment using the atomic force microscope (AFM) stands as a unique method. AFM directly measures the resistance from unfolding, which can be quantified from the observed force-extension profile. It has been shown that key features observed in an AFM pulling experiment can be well reproduced by computational molecular dynamics simulations. Here, we applied computational pulling for estimating the accuracy of computational protein structure models under the hypothesis that the structural stability would positively correlated with the accuracy, i.e. the closeness to the native, of a model. We used in total 4929 structure models for 24 target proteins from the Critical Assessment of Techniques of Structure Prediction (CASP) and investigated if the magnitude of the break force, that is, the force required to rearrange the model's structure, from the force profile was sufficient information for selecting near-native models. We found that near-native models can be successfully selected by examining their break forces suggesting that high break force indeed indicates high stability of models. On the other hand, there were also near-native models that had relatively low peak forces. The mechanisms of the stability exhibited by the break forces were explored and discussed.

TRScore: a 3D RepVGG-based scoring method for ranking protein docking models.

Protein-protein interactions (PPI) play important roles in cellular activities. Due to the technical difficulty and high cost of experimental methods, there are considerable interests towards the development of computational approaches, such as protein docking, to decipher PPI patterns. One of the important and difficult aspects in protein docking is recognizing near-native conformations from a set of decoys, but unfortunately, traditional scoring functions still suffer from limited accuracy. Therefore, new scoring methods are pressingly needed in methodological and/or practical implications. We present a new deep learning-based scoring method for ranking protein-protein docking models based on a 3D RepVGG network, named TRScore. To recognize near-native conformations from a set of decoys, TRScore voxelizes the protein-protein interface into a 3D grid labeled by the number of atoms in different physicochemical classes. Benefiting from the deep convolutional RepVGG architecture, TRScore can effectively capture the subtle differences between energetically favorable near-native models and unfavorable non-native decoys without needing extra information. TRScore was extensively evaluated on diverse test sets including protein-protein docking benchmark 5.0 update set, DockGround decoy set, as well as realistic CAPRI decoy set and overall obtained a significant improvement over existing methods in cross-validation and independent evaluations. Codes available at: https://github.com/BioinformaticsCSU/TRScore.

Quality Assessment of Protein Docking Models Based on Graph Neural Network.

Protein docking provides a structural basis for the design of drugs and vaccines. Among the processes of protein docking, quality assessment (QA) is utilized to pick near-native models from numerous protein docking candidate conformations, and it directly determines the final docking results. Although extensive efforts have been made to improve QA accuracy, it is still the bottleneck of current protein docking systems. In this paper, we presented a Deep Graph Attention Neural Network (DGANN) to evaluate and rank protein docking candidate models. DGANN learns inter-residue physio-chemical properties and structural fitness across the two protein monomers in a docking model and generates their probabilities of near-native models. On the ZDOCK decoy benchmark, our DGANN outperformed the ranking provided by ZDOCK in terms of ranking good models into the top selections. Furthermore, we conducted comparative experiments on an independent testing dataset, and the results also demonstrated the superiority and generalization of our proposed method.

Protein Structure Prediction from NMR Hydrogen-Deuterium Exchange Data.

Amide hydrogen-deuterium exchange (HDX) has long been used to determine regional flexibility and binding sites in proteins; however, the data are too sparse for full structural characterization. Experiments that measure HDX rates, such as HDX-NMR, have far higher throughput compared to structure determination via X-ray crystallography, cryo-EM, or a full suite of NMR experiments. Data from HDX-NMR experiments encode information on the protein structure, making HDX a prime candidate to be supplemented by computational algorithms for protein structure prediction. We have developed a methodology to incorporate HDX-NMR data into ab initio protein structure prediction using the Rosetta software framework to predict structures based on experimental agreement. To demonstrate the efficacy of our algorithm, we examined 38 proteins with HDX-NMR data available, comparing the predicted model with and without the incorporation of HDX data into scoring. The root-mean-square deviation (rmsd, a measure of the average atomic distance between superimposed models) of the predicted model improved by 1.42 Å on average after incorporating the HDX-NMR data into scoring. The average rmsd improvement for the proteins where the selected model rmsd changed after incorporating HDX data was 3.63 Å, including one improvement of more than 11 Å and seven proteins improving by greater than 4 Å, with 12/15 proteins improving overall. Additionally, for independent verification, two proteins that were not part of the original benchmark were scored including HDX data, with a dramatic improvement of the selected model rmsd of nearly 9 Å for one of the proteins. Moreover, we have developed a confidence metric allowing us to successfully identify near-native models in the absence of a native structure. Improvement in model selection with a strong confidence measure demonstrates that protein structure prediction with HDX-NMR is a powerful tool which can be performed with minimal additional computational strain and expense.

Covalent flexible peptide docking in Rosetta.

Electrophilic peptides that form an irreversible covalent bond with their target have great potential for binding targets that have been previously considered undruggable. However, the discovery of such peptides remains a challenge. Here, we present Rosetta CovPepDock, a computational pipeline for peptide docking that incorporates covalent binding between the peptide and a receptor cysteine. We applied CovPepDock retrospectively to a dataset of 115 disulfide-bound peptides and a dataset of 54 electrophilic peptides. It produced a top-five scoring, near-native model, in 89% and 100% of the cases when docking from the native conformation, and 20% and 90% when docking from an extended peptide conformation, respectively. In addition, we developed a protocol for designing electrophilic peptide binders based on known non-covalent binders or protein–protein interfaces. We identified 7154 peptide candidates in the PDB for application of this protocol. As a proof-of-concept we validated the protocol on the non-covalent complex of 14-3-3σ and YAP1 phosphopeptide. The protocol identified seven highly potent and selective irreversible peptide binders. The predicted binding mode of one of the peptides was validated using X-ray crystallography. This case-study demonstrates the utility and impact of CovPepDock. It suggests that many new electrophilic peptide binders can be rapidly discovered, with significant potential as therapeutic molecules and chemical probes.

Novel sampling strategies and a coarse-grained score function for docking homomers, flexible heteromers, and oligosaccharides using Rosetta in CAPRI rounds 37-45.

Critical Assessment of PRediction of Interactions (CAPRI) rounds 37 through 45 introduced larger complexes, new macromolecules, and multistage assemblies. For these rounds, we used and expanded docking methods in Rosetta to model 23 target complexes. We successfully predicted 14 target complexes and recognized and refined near-native models generated by other groups for two further targets. Notably, for targets T110 and T136, we achieved the closest prediction of any CAPRI participant. We created several innovative approaches during these rounds. Since round 39 (target 122), we have used the new RosettaDock 4.0, which has a revamped coarse-grained energy function and the ability to perform conformer selection during docking with hundreds of pregenerated protein backbones. Ten of the complexes had some degree of symmetry in their interactions, so we tested Rosetta SymDock, realized its shortcomings, and developed the next-generation symmetric docking protocol, SymDock2, which includes docking of multiple backbones and induced-fit refinement. Since the last CAPRI assessment, we also developed methods for modeling and designing carbohydrates in Rosetta, and we used them to successfully model oligosaccharide-protein complexes in round 41. Although the results were broadly encouraging, they also highlighted the pressing need to invest in (a) flexible docking algorithms with the ability to model loop and linker motions and in (b) new sampling and scoring methods for oligosaccharide-protein interactions.

Protein docking model evaluation by 3D deep convolutional neural networks.

Many important cellular processes involve physical interactions of proteins. Therefore, determining protein quaternary structures provide critical insights for understanding molecular mechanisms of functions of the complexes. To complement experimental methods, many computational methods have been developed to predict structures of protein complexes. One of the challenges in computational protein complex structure prediction is to identify near-native models from a large pool of generated models. We developed a convolutional deep neural network-based approach named DOcking decoy selection with Voxel-based deep neural nEtwork (DOVE) for evaluating protein docking models. To evaluate a protein docking model, DOVE scans the protein-protein interface of the model with a 3D voxel and considers atomic interaction types and their energetic contributions as input features applied to the neural network. The deep learning models were trained and validated on docking models available in the ZDock and DockGround databases. Among the different combinations of features tested, almost all outperformed existing scoring functions. Codes available at http://github.com/kiharalab/DOVE, http://kiharalab.org/dove/. Supplementary data are available at Bioinformatics online.

Refined template selection and combination algorithm significantly improves template-based modeling accuracy.

In contrast to ab-initio protein modeling methodologies, comparative modeling is considered as the most popular and reliable algorithm to model protein structure. However, the selection of the best set of templates is still a major challenge. An effective template-ranking algorithm is developed to efficiently select only the reliable hits for predicting the protein structures. The algorithm employs the pairwise as well as multiple sequence alignments of template hits to rank and select the best possible set of templates. It captures several key sequences and structural information of template hits and converts into scores to effectively rank them. This selected set of templates is used to model a target. Modeling accuracy of the algorithm is tested and evaluated on TBM-HA domain containing CASP8, CASP9 and CASP10 targets. On an average, this template ranking and selection algorithm improves GDT-TS, GDT-HA and TM_Score by 3.531, 4.814 and 0.022, respectively. Further, it has been shown that the inclusion of structurally similar templates with ample conformational diversity is crucial for the modeling algorithm to maximally as well as reliably span the target sequence and construct its near-native model. The optimal model sampling also holds the key to predict the best possible target structure.

More efficient screening of protein-protein complex model structures for reducing the number of candidates.

Rigid-body protein-protein docking is very efficient in generating tens of thousands of docked complex models (decoys) in a very short time without considering structure change upon binding, but typical docking scoring functions are not necessarily sufficiently accurate to narrow these decoys down to a small number of plausible candidates. Flexible refinements and sophisticated evaluation of the decoys are thus required to achieve more accurate prediction. Since this process is time-consuming, an efficient screening method to reduce the number of decoys is necessary immediately following rigid-body dockings. We attempted to develop an efficient screening method by clustering decoys generated by the rigid-body docking ZDOCK. We introduced the three metrics ligand-root-mean-square deviation (L-RMSD), interface-ligand-RMSD (iL-RMSD), and the fraction of common contacts (FCC), and examined various ranges of cut-offs for clusters to determine the best set of clustering parameters. Although the employed clustering algorithm is simple, it successfully reduced the number of decoys. Using iL-RMSD with a cut-off radius of 8 Å, the number of decoys that contain at least one near-native model with 90% probability decreased from 4,808 to 320, a 93% reduction in the original number of decoys. Using FCC for the clustering step, the top 1,000 success rates, defined as the probability that the top 1,000 models contain at least one near-native structure, reached 97%. We conclude that the proposed method is very efficient in selecting a small number of decoys that include near-native decoys.

An SVM-based method for assessment of transcription factor-DNA complex models

BackgroundAtomic details of protein-DNA complexes can provide insightful information for better understanding of the function and binding specificity of DNA binding proteins. In addition to experimental methods for solving protein-DNA complex structures, protein-DNA docking can be used to predict native or near-native complex models. A docking program typically generates a large number of complex conformations and predicts the complex model(s) based on interaction energies between protein and DNA. However, the prediction accuracy is hampered by current approaches to model assessment, especially when docking simulations fail to produce any near-native models.ResultsWe present here a Support Vector Machine (SVM)-based approach for quality assessment of the predicted transcription factor (TF)-DNA complex models. Besides a knowledge-based protein-DNA interaction potential DDNA3, we applied several structural features that have been shown to play important roles in binding specificity between transcription factors and DNA molecules to quality assessment of complex models. To address the issue of unbalanced positive and negative cases in the training dataset, we applied hard-negative mining, an iterative training process that selects an initial training dataset by combining all of the positive cases and a random sample from the negative cases. Results show that the SVM model greatly improves prediction accuracy (84.2%) over two knowledge-based protein-DNA interaction potentials, orientation potential (60.8%) and DDNA3 (68.4%). The improvement is achieved through reducing the number of false positive predictions, especially for the hard docking cases, in which a docking algorithm fails to produce any near-native complex models.ConclusionsA learning-based SVM scoring model with structural features for specific protein-DNA binding and an atomic-level protein-DNA interaction potential DDNA3 significantly improves prediction accuracy of complex models by successfully identifying cases without near-native structural models.

Selecting near‐native protein structures from ab initio models using ensemble clustering

BackgroundAb initio protein structure prediction is to predict the tertiary structure of a protein from its amino acid sequence alone. As an important topic in bioinformatics, considerable efforts have been made on designing the ab initio methods. Unfortunately, lacking of a perfect energy function, it is a difficult task to select a good near‐native structure from the predicted decoy structures in the last step.MethodsHere we propose an ensemble clustering method based on k‐medoids to deal with this problem. The k‐medoids method is run many times to generate clustering ensembles, and then a voting method is used to combine the clustering results. A confidence score is defined to select the final near‐native model, considering both the cluster size and the cluster similarity.ResultsWe have applied the method to 54 single‐domain targets in CASP‐11. For about 70.4% of these targets, the proposed method can select better near‐native structures compared to the SPICKER method used by the I‐TASSER server.ConclusionsThe experiments show that, the proposed method is effective in selecting the near‐native structure from decoy sets for different targets in terms of the similarity between the selected structure and the native structure.

De novo computational RNA modeling into cryo-EM maps of large ribonucleoprotein complexes.

Increasingly, cryo-electron microscopy (cryo-EM) is used to determine the structures of RNA-protein assemblies, but nearly all maps determined with this method have biologically important regions where the local resolution does not permit RNA coordinate tracing. To address these omissions, we present de novo ribonucleoprotein modeling in real space through assembly of fragments together with experimental density in Rosetta (DRRAFTER). We show that DRRAFTER recovers near-native models for a diverse benchmark set of RNA-protein complexes including the spliceosome, mitochondrial ribosome, and CRISPR-Cas9-sgRNA complexes; rigorous blind tests include yeast U1 snRNP and spliceosomal P complex maps. Additionally, to aid in model interpretation, we present a method for reliable in situ estimation of DRRAFTER model accuracy. Finally, we apply DRRAFTER to recently determined maps of telomerase, the HIV-1 reverse transcriptase initiation complex, and the packaged MS2 genome, demonstrating the acceleration of accurate model building in challenging cases.

Protein docking refinement by convex underestimation in the low-dimensional subspace of encounter complexes

We propose a novel stochastic global optimization algorithm with applications to the refinement stage of protein docking prediction methods. Our approach can process conformations sampled from multiple clusters, each roughly corresponding to a different binding energy funnel. These clusters are obtained using a density-based clustering method. In each cluster, we identify a smooth “permissive” subspace which avoids high-energy barriers and then underestimate the binding energy function using general convex polynomials in this subspace. We use the underestimator to bias sampling towards its global minimum. Sampling and subspace underestimation are repeated several times and the conformations sampled at the last iteration form a refined ensemble. We report computational results on a comprehensive benchmark of 224 protein complexes, establishing that our refined ensemble significantly improves the quality of the conformations of the original set given to the algorithm. We also devise a method to enhance the ensemble from which near-native models are selected.

Natural language processing in text mining for structural modeling of protein complexes

BackgroundStructural modeling of protein-protein interactions produces a large number of putative configurations of the protein complexes. Identification of the near-native models among them is a serious challenge. Publicly available results of biomedical research may provide constraints on the binding mode, which can be essential for the docking. Our text-mining (TM) tool, which extracts binding site residues from the PubMed abstracts, was successfully applied to protein docking (Badal et al., PLoS Comput Biol, 2015; 11: e1004630). Still, many extracted residues were not relevant to the docking.ResultsWe present an extension of the TM tool, which utilizes natural language processing (NLP) for analyzing the context of the residue occurrence. The procedure was tested using generic and specialized dictionaries. The results showed that the keyword dictionaries designed for identification of protein interactions are not adequate for the TM prediction of the binding mode. However, our dictionary designed to distinguish keywords relevant to the protein binding sites led to considerable improvement in the TM performance. We investigated the utility of several methods of context analysis, based on dissection of the sentence parse trees. The machine learning-based NLP filtered the pool of the mined residues significantly more efficiently than the rule-based NLP. Constraints generated by NLP were tested in docking of unbound proteins from the DOCKGROUND X-ray benchmark set 4. The output of the global low-resolution docking scan was post-processed, separately, by constraints from the basic TM, constraints re-ranked by NLP, and the reference constraints. The quality of a match was assessed by the interface root-mean-square deviation. The results showed significant improvement of the docking output when using the constraints generated by the advanced TM with NLP.ConclusionsThe basic TM procedure for extracting protein-protein binding site residues from the PubMed abstracts was significantly advanced by the deep parsing (NLP techniques for contextual analysis) in purging of the initial pool of the extracted residues. Benchmarking showed a substantial increase of the docking success rate based on the constraints generated by the advanced TM with NLP.

Comparison of Full and Interface Structure Alignment in Template-Based Protein Docking

Template-based protein docking, based on sequence/structure similarity to co-crystalyzed complexes complements ab initio docking in predicting the structure of protein-protein complexes. When highly homologous templates for the entire target complex are not available, docking by interface or partial structure alignment (PSA) may provide an adequate solution. Earlier studies indicated that PSA docking can also be used to further refine docking models generated by the full structure alignment (FSA) of the target and the template. We present a thorough comparative analysis of FSA and PSA docking methodologies based on the docking predictions for bound and unbound proteins from the Dockground benchmark set 4 (http://dockground.compbio.ku.edu). Using the unbound structures in docking (i.e. accounting for conformational changes upon binding) decreases the docking performance only slightly for both methods, compared to the docking of the bound structures (by ∼0.3% in top 10 predictions). FSA selects better templates for proteins with different function, whereas PSA provides better templates for multi-domain proteins in which only one domain participates in the interaction (e.g., enzyme-inhibitor). Template structures that have matching CATH annotations with the target yield better docking performance (the success rate 65% and 67% in top 100, for FSA and PSA, respectively, as opposed to 44% and 46% without the CATH annotation). Significant improvement in the unbound docking success rate (from ∼30 to 60% in top 10) is observed when the same template is selected by both PSA and FSA. By adding templates identified by sophisticated sequence/evolutionary based methods, we were able to build near-native models for 90% of complexes in our target set.