In the field of medical imaging analysis, particularly in interpreting chest X-rays, deep learning models have shown remarkable progress. Nonetheless, these models often face challenges such as limited annotation and inadequate utilization of public data resources. This is particularly apparent with databases containing multimodal data, such as images and medical reports, where the effective integration of this multimodal information remains difficult. To address these limitations, we propose the Neighbor-Assisted Multimodal Attention Network (NAMAN), a novel approach designed to leverage retrieval augmentation techniques to enhance disease classification performance. NAMAN combines nearest neighbor search with multimodal fusion, utilizing both visual features from similar X-ray images and textual information from corresponding medical records. The experimental results demonstrate the efficacy of incorporating retrieved neighbor information and multimodal integration mechanisms in NAMAN. Our ablation studies offer insights into the optimal configuration of the model, including the effects of various attention mechanisms and the number of retrieved neighbors. This work contributes to the expanding field of retrieval-augmented approaches in medical imaging, presenting a promising avenue for leveraging large-scale, multimodal medical databases to enhance diagnostic accuracy and reliability.

With the rapid progress of deep learning in cryo-electron microscopy and protein structure prediction, improving the accuracy of the protein structure model by using a density map and predicted contact/distance map through deep learning has become an urgent need for robust methods. Thus, designing an effective protein structure optimization strategy based on the density map and predicted contact/distance map is critical to improving the accuracy of structure refinement. In this article, a protein structure optimization method based on the density map and predicted contact/distance map by deep-learning technology was proposed in accordance with the result of matching between the density map and the initial model. Physics- and knowledge-based energy functions, integrated with Cryo-EM density map data and deep-learning data, were used to optimize the protein structure in the simulation. The dynamic confidence score was introduced to the iterative process for choosing whether it is a density map or a contact/distance map to dominate the movement in the simulation to improve the accuracy of refinement. The protocol was tested on a large set of 224 non-homologous membrane proteins and generated 214 structural models with correct folds, where 4.5% of structural models were generated from structural models with incorrect folds. Compared with other state-of-the-art methods, the major advantage of the proposed methods lies in the skills for using density map and contact/distance map in the simulation, as well as the new energy function in the re-assembly simulations. Overall, the results demonstrated that this strategy is a valuable approach and ready to use for atomic-level structure refinement using cryo-EM density map and predicted contact/distance map.

MOTIVATION: Accurate and efficient predictions of protein structures play an important role in understanding their functions. Iterative Threading Assembly Refinement (I-TASSER) is one of the most successful and widely used protein structure prediction methods in the recent community-wide CASP experiments. Yet, the computational efficiency of I-TASSER is one of the limiting factors that prevent its application for large-scale structure modeling. RESULTS: We present I-TASSER for Graphics Processing Units (GPU-I-TASSER), a GPU accelerated I-TASSER protein structure prediction tool for fast and accurate protein structure prediction. Our implementation is based on OpenACC parallelization of the replica-exchange Monte Carlo simulations to enhance the speed of I-TASSER by extending its capabilities to the GPU architecture. On a benchmark dataset of 71 protein structures, GPU-I-TASSER achieves on average a 10× speedup with comparable structure prediction accuracy compared to the CPU version of the I-TASSER. AVAILABILITY AND IMPLEMENTATION: The complete source code for GPU-I-TASSER can be downloaded and used without restriction from https://zhanggroup.org/GPU-I-TASSER/. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Despite the immense progress recently witnessed in protein structure prediction, the modeling accuracy for proteins that lack sequence and/or structure homologs remains to be improved. We developed an open-source program, DeepFold, which integrates spatial restraints predicted by multi-task deep residual neural-networks along with a knowledge-based energy function to guide its gradient-descent folding simulations. The results on large-scale benchmark tests showed that DeepFold creates full-length models with accuracy significantly beyond classical folding approaches and other leading deep learning methods. Of particular interest is the modeling performance on the most difficult targets with very few homologous sequences, where DeepFold achieved an average TM-score that was 40.3% higher than trRosetta and 44.9% higher than DMPfold. Furthermore, the folding simulations for DeepFold were 262 times faster than traditional fragment assembly simulations. These results demonstrate the power of accurately predicted deep learning potentials to improve both the accuracy and speed of ab initio protein structure prediction.

Accurate identification of protein function is critical to elucidate life mechanisms and design new drugs. We proposed a novel deep-learning method, ATGO, to predict Gene Ontology (GO) attributes of proteins through a triplet neural-network architecture embedded with pre-trained language models from protein sequences. The method was systematically tested on 1068 non-redundant benchmarking proteins and 3328 targets from the third Critical Assessment of Protein Function Annotation (CAFA) challenge. Experimental results showed that ATGO achieved a significant increase of the GO prediction accuracy compared to the state-of-the-art approaches in all aspects of molecular function, biological process, and cellular component. Detailed data analyses showed that the major advantage of ATGO lies in the utilization of pre-trained transformer language models which can extract discriminative functional pattern from the feature embeddings. Meanwhile, the proposed triplet network helps enhance the association of functional similarity with feature similarity in the sequence embedding space. In addition, it was found that the combination of the network scores with the complementary homology-based inferences could further improve the accuracy of the predicted models. These results demonstrated a new avenue for high-accuracy deep-learning function prediction that is applicable to large-scale protein function annotations from sequence alone.

X-ray crystallography is the major approach for determining atomic-level protein structures. Because not all proteins can be easily crystallized, accurate prediction of protein crystallization propensity provides critical help in guiding experimental design and improving the success rate of X-ray crystallography experiments. This study has developed a new machine-learning-based pipeline that uses a newly developed deep-cascade forest (DCF) model with multiple types of sequence-based features to predict protein crystallization propensity. Based on the developed pipeline, two new protein crystallization propensity predictors, denoted as DCFCrystal and MDCFCrystal, have been implemented. DCFCrystal is a multistage predictor that can estimate the success propensities of the three individual steps (production of protein material, purification and production of crystals) in the protein crystallization process. MDCFCrystal is a single-stage predictor that aims to estimate the probability that a protein will pass through the entire crystallization process. Moreover, DCFCrystal is designed for general proteins, whereas MDCFCrystal is specially designed for membrane proteins, which are notoriously difficult to crystalize. DCFCrystal and MDCFCrystal were separately tested on two benchmark datasets consisting of 12 289 and 950 proteins, respectively, with known crystallization results from various experimental records. The experimental results demonstrated that DCFCrystal and MDCFCrystal increased the value of Matthew's correlation coefficient by 199.7% and 77.8%, respectively, compared to the best of other state-of-the-art protein crystallization propensity predictors. Detailed analyses show that the major advantages of DCFCrystal and MDCFCrystal lie in the efficiency of the DCF model and the sensitivity of the sequence-based features used, especially the newly designed pseudo-predicted hybrid solvent accessibility (PsePHSA) feature, which improves crystallization recognition by incorporating sequence-order information with solvent accessibility of residues. Meanwhile, the new crystal-dataset constructions help to train the models with more comprehensive crystallization knowledge.

MOTIVATION: The mathematically optimal solution in computational protein folding simulations does not always correspond to the native structure, due to the imperfection of the energy force fields. There is therefore a need to search for more diverse suboptimal solutions in order to identify the states close to the native. We propose a novel multimodal optimization protocol to improve the conformation sampling efficiency and modeling accuracy of de novo protein structure folding simulations. RESULTS: A distance-assisted multimodal optimization sampling algorithm, MMpred, is proposed for de novo protein structure prediction. The protocol consists of three stages: The first is a modal exploration stage, in which a structural similarity evaluation model DMscore is designed to control the diversity of conformations, generating a population of diverse structures in different low-energy basins. The second is a modal maintaining stage, where an adaptive clustering algorithm MNDcluster is proposed to divide the populations and merge the modal by adjusting the annealing temperature to locate the promising basins. In the last stage of modal exploitation, a greedy search strategy is used to accelerate the convergence of the modal. Distance constraint information is used to construct the conformation scoring model to guide sampling. MMpred is tested on a large set of 320 non-redundant proteins, where MMpred obtains models with TM-score≥0.5 on 291 cases, which is 28% higher than that of Rosetta guided with the same set of distance constraints. In addition, on 320 benchmark proteins, the enhanced version of MMpred (E-MMpred) has 167 targets better than trRosetta when the best of five models are evaluated. The average TM-score of the best model of E-MMpred is 0.732, which is comparable to trRosetta (0.730). AVAILABILITY AND IMPLEMENTATION: The source code and executable are freely available at https://github.com/iobio-zjut/MMpred. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

The topology of protein folds can be specified by the inter-residue contact-maps and accurate contact-map prediction can help ab initio structure folding. We developed TripletRes to deduce protein contact-maps from discretized distance profiles by end-to-end training of deep residual neural-networks. Compared to previous approaches, the major advantage of TripletRes is in its ability to learn and directly fuse a triplet of coevolutionary matrices extracted from the whole-genome and metagenome databases and therefore minimize the information loss during the course of contact model training. TripletRes was tested on a large set of 245 non-homologous proteins from CASP 11&12 and CAMEO experiments and outperformed other top methods from CASP12 by at least 58.4% for the CASP 11&12 targets and 44.4% for the CAMEO targets in the top-L long-range contact precision. On the 31 FM targets from the latest CASP13 challenge, TripletRes achieved the highest precision (71.6%) for the top-L/5 long-range contact predictions. It was also shown that a simple re-training of the TripletRes model with more proteins can lead to further improvement with precisions comparable to state-of-the-art methods developed after CASP13. These results demonstrate a novel efficient approach to extend the power of deep convolutional networks for high-accuracy medium- and long-range protein contact-map predictions starting from primary sequences, which are critical for constructing 3D structure of proteins that lack homologous templates in the PDB library.

Three-dimensional structures of proteins can provide important clues into the efficacy of personalized treatment. We perform a structural analysis of variants within three inherited lysosomal storage disorders, comparing variants responsive to pharmacological chaperone treatment to those unresponsive to such treatment. We find that predicted ΔΔG of mutation is higher on average for variants unresponsive to treatment, in the case of datasets for both Fabry disease and Pompe disease, in line with previous findings. Using both a single decision tree and an advanced machine learning approach based on the larger Fabry dataset, we correctly predict responsiveness of three Gaucher disease variants, and we provide predictions for untested variants. Many variants are predicted to be responsive to treatment, suggesting that drug-based treatments may be effective for a number of variants in Gaucher disease. In our analysis, we observe dependence on a topological feature reporting on contact arrangements which is likely connected to the order of folding of protein residues, and we provide a potential justification for this observation based on steady-state cellular kinetics.

MOTIVATION: Protein structure and function are essentially determined by how the side-chain atoms interact with each other. Thus, accurate protein side-chain packing (PSCP) is a critical step toward protein structure prediction and protein design. Despite the importance of the problem, however, the accuracy and speed of current PSCP programs are still not satisfactory. RESULTS: We present FASPR for fast and accurate PSCP by using an optimized scoring function in combination with a deterministic searching algorithm. The performance of FASPR was compared with four state-of-the-art PSCP methods (CISRR, RASP, SCATD and SCWRL4) on both native and non-native protein backbones. For the assessment on native backbones, FASPR achieved a good performance by correctly predicting 69.1% of all the side-chain dihedral angles using a stringent tolerance criterion of 20°, compared favorably with SCWRL4, CISRR, RASP and SCATD which successfully predicted 68.8%, 68.6%, 67.8% and 61.7%, respectively. Additionally, FASPR achieved the highest speed for packing the 379 test protein structures in only 34.3 s, which was significantly faster than the control methods. For the assessment on non-native backbones, FASPR showed an equivalent or better performance on I-TASSER predicted backbones and the backbones perturbed from experimental structures. Detailed analyses showed that the major advantage of FASPR lies in the optimal combination of the dead-end elimination and tree decomposition with a well optimized scoring function, which makes FASPR of practical use for both protein structure modeling and protein design studies. AVAILABILITY AND IMPLEMENTATION: The web server, source code and datasets are freely available at https://zhanglab.ccmb.med.umich.edu/FASPR and https://github.com/tommyhuangthu/FASPR. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: The accuracy and success rate of de novo protein design remain limited, mainly due to the parameter over-fitting of current energy functions and their inability to discriminate incorrect designs from correct designs. RESULTS: We developed an extended energy function, EvoEF2, for efficient de novo protein sequence design, based on a previously proposed physical energy function, EvoEF. Remarkably, EvoEF2 recovered 32.5%, 47.9% and 22.3% of all, core and surface residues for 148 test monomers, and was generally applicable to protein-protein interaction design, as it recapitulated 30.9%, 42.4%, 31.3% and 21.4% of all, core, interface and surface residues for 88 test dimers, significantly outperforming EvoEF on the native sequence recapitulation. We further used I-TASSER to evaluate the foldability of the 148 designed monomer sequences, where all of them were predicted to fold into structures with high fold- and atomic-level similarity to their corresponding native structures, as demonstrated by the fact that 87.8% of the predicted structures shared a root-mean-square-deviation less than 2 Å to their native counterparts. The study also demonstrated that the usefulness of physical energy functions is highly correlated with the parameter optimization processes, and EvoEF2, with parameters optimized using sequence recapitulation, is more suitable for computational protein sequence design than EvoEF, which was optimized on thermodynamic mutation data. AVAILABILITY AND IMPLEMENTATION: The source code of EvoEF2 and the benchmark datasets are freely available at https://zhanglab.ccmb.med.umich.edu/EvoEF. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: Most proteins perform their biological functions through interactions with other proteins in cells. Amino acid mutations, especially those occurring at protein interfaces, can change the stability of protein-protein interactions (PPIs) and impact their functions, which may cause various human diseases. Quantitative estimation of the binding affinity changes (ΔΔGbind) caused by mutations can provide critical information for protein function annotation and genetic disease diagnoses. RESULTS: We present SSIPe, which combines protein interface profiles, collected from structural and sequence homology searches, with a physics-based energy function for accurate ΔΔGbind estimation. To offset the statistical limits of the PPI structure and sequence databases, amino acid-specific pseudocounts were introduced to enhance the profile accuracy. SSIPe was evaluated on large-scale experimental data containing 2204 mutations from 177 proteins, where training and test datasets were stringently separated with the sequence identity between proteins from the two datasets below 30%. The Pearson correlation coefficient between estimated and experimental ΔΔGbind was 0.61 with a root-mean-square-error of 1.93 kcal/mol, which was significantly better than the other methods. Detailed data analyses revealed that the major advantage of SSIPe over other traditional approaches lies in the novel combination of the physical energy function with the new knowledge-based interface profile. SSIPe also considerably outperformed a former profile-based method (BindProfX) due to the newly introduced sequence profiles and optimized pseudocount technique that allows for consideration of amino acid-specific prior mutation probabilities. AVAILABILITY AND IMPLEMENTATION: Web-server/standalone program, source code and datasets are freely available at https://zhanglab.ccmb.med.umich.edu/SSIPe and https://github.com/tommyhuangthu/SSIPe. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: Regions that connect secondary structure elements in a protein are known as loops, whose slight change will produce dramatic effect on the entire topology. This study investigates whether the accuracy of protein structure prediction can be improved using a loop-specific sampling strategy. RESULTS: A novel de novo protein structure prediction method that combines global exploration and loop perturbation is proposed in this study. In the global exploration phase, the fragment recombination and assembly are used to explore the massive conformational space and generate native-like topology. In the loop perturbation phase, a loop-specific local perturbation model is designed to improve the accuracy of the conformation and is solved by differential evolution algorithm. These two phases enable a cooperation between global exploration and local exploitation. The filtered contact information is used to construct the conformation selection model for guiding the sampling. The proposed CGLFold is tested on 145 benchmark proteins, 14 free modeling (FM) targets of CASP13 and 29 FM targets of CASP12. The experimental results show that the loop-specific local perturbation can increase the structure diversity and success rate of conformational update and gradually improve conformation accuracy. CGLFold obtains template modeling score ≥ 0.5 models on 95 standard test proteins, 7 FM targets of CASP13 and 9 FM targets of CASP12. AVAILABILITY AND IMPLEMENTATION: The source code and executable versions are freely available at https://github.com/iobio-zjut/CGLFold. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: Template-based and template-free methods have both been widely used in predicting the structures of protein-protein complexes. Template-based modeling is effective when a reliable template is available, while template-free methods are required for predicting the binding modes or interfaces that have not been previously observed. Our goal is to combine the two methods to improve computational protein-protein complex structure prediction. RESULTS: Here, we present a method to identify and combine high-confidence predictions of a template-based method (SPRING) with a template-free method (ZDOCK). Cross-validated using the protein-protein docking benchmark version 5.0, our method (ZING) achieved a success rate of 68.2%, outperforming SPRING and ZDOCK, with success rates of 52.1% and 35.9% respectively, when the top 10 predictions were considered per test case. In conclusion, a statistics-based method that evaluates and integrates predictions from template-based and template-free methods is more successful than either method independently. AVAILABILITY AND IMPLEMENTATION: ZING is available for download as a Github repository (https://github.com/weng-lab/ZING.git). SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: Protein structure refinement is an important step of protein structure prediction. Existing approaches have generally used a single scoring function combined with Monte Carlo method or Molecular Dynamics algorithm. The one-dimension optimization of a single energy function may take the structure too far away without a constraint. The basic motivation of our study is to reduce the bias problem caused by minimizing only a single energy function due to the very diversity of different protein structures. RESULTS: We report a new Artificial Intelligence-based protein structure Refinement method called AIR. Its fundamental idea is to use multiple energy functions as multi-objectives in an effort to correct the potential inaccuracy from a single function. A multi-objective particle swarm optimization algorithm-based structure refinement is designed, where each structure is considered as a particle in the protocol. With the refinement iterations, the particles move around. The quality of particles in each iteration is evaluated by three energy functions, and the non-dominated particles are put into a set called Pareto set. After enough iteration times, particles from the Pareto set are screened and part of the top solutions are outputted as the final refined structures. The multi-objective energy function optimization strategy designed in the AIR protocol provides a different constraint view of the structure, by extending the one-dimension optimization to a new three-dimension space optimization driven by the multi-objective particle swarm optimization engine. Experimental results on CASP11, CASP12 refinement targets and blind tests in CASP 13 turn to be promising. AVAILABILITY AND IMPLEMENTATION: The AIR is available online at: www.csbio.sjtu.edu.cn/bioinf/AIR/. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: Many protein function databases are built on automated or semi-automated curations and can contain various annotation errors. The correction of such misannotations is critical to improving the accuracy and reliability of the databases. RESULTS: We proposed a new approach to detect potentially incorrect Gene Ontology (GO) annotations by comparing the ratio of annotation rates (RAR) for the same GO term across different taxonomic groups, where those with a relatively low RAR usually correspond to incorrect annotations. As an illustration, we applied the approach to 20 commonly studied species in two recent UniProt-GOA releases and identified 250 potential misannotations in the 2018-11-6 release, where only 25% of them were corrected in the 2019-6-3 release. Importantly, 56% of the misannotations are 'Inferred from Biological aspect of Ancestor (IBA)' which is in contradiction with previous observations that attributed misannotations mainly to 'Inferred from Sequence or structural Similarity (ISS)', probably reflecting an error source shift due to the new developments of function annotation databases. The results demonstrated a simple but efficient misannotation detection approach that is useful for large-scale comparative protein function studies. AVAILABILITY AND IMPLEMENTATION: https://zhanglab.ccmb.med.umich.edu/RAR. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

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.

MOTIVATION: Protein domains are subunits that can fold and function independently. Correct domain boundary assignment is thus a critical step toward accurate protein structure and function analyses. There is, however, no efficient algorithm available for accurate domain prediction from sequence. The problem is particularly challenging for proteins with discontinuous domains, which consist of domain segments that are separated along the sequence. RESULTS: We developed a new algorithm, FUpred, which predicts protein domain boundaries utilizing contact maps created by deep residual neural networks coupled with coevolutionary precision matrices. The core idea of the algorithm is to retrieve domain boundary locations by maximizing the number of intra-domain contacts, while minimizing the number of inter-domain contacts from the contact maps. FUpred was tested on a large-scale dataset consisting of 2549 proteins and generated correct single- and multi-domain classifications with a Matthew's correlation coefficient of 0.799, which was 19.1% (or 5.3%) higher than the best machine learning (or threading)-based method. For proteins with discontinuous domains, the domain boundary detection and normalized domain overlapping scores of FUpred were 0.788 and 0.521, respectively, which were 17.3% and 23.8% higher than the best control method. The results demonstrate a new avenue to accurately detect domain composition from sequence alone, especially for discontinuous, multi-domain proteins. AVAILABILITY AND IMPLEMENTATION: https://zhanglab.ccmb.med.umich.edu/FUpred. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Various mutation strategies show distinct advantages in differential evolution (DE). The cooperation of multiple strategies in the evolutionary process may be effective. This paper presents an underestimation-assisted global and local cooperative DE to simultaneously enhance the effectiveness and efficiency. In the proposed algorithm, two phases, namely, the global exploration and the local exploitation, are performed in each generation. In the global phase, a set of trial vectors is produced for each target individual by employing multiple strategies with strong exploration capability. Afterward, an adaptive underestimation model with a self-adapted slope control parameter is proposed to evaluate these trial vectors, the best of which is selected as the candidate. In the local phase, the better-based strategies guided by individuals that are better than the target individual are designed. For each individual accepted in the global phase, multiple trial vectors are generated by using these strategies and filtered by the underestimation value. The cooperation between the global and local phases includes two aspects. First, both of them concentrate on generating better individuals for the next generation. Second, the global phase aims to locate promising regions quickly while the local phase serves as a local search for enhancing convergence. Moreover, a simple mechanism is designed to determine the parameter of DE adaptively in the searching process. Finally, the proposed approach is applied to predict the protein 3D structure. Experimental studies on classical benchmark functions, CEC test sets, and protein structure prediction problem show that the proposed approach is superior to the competitors.

MOTIVATION: Comparison of RNA 3D structures can be used to infer functional relationship of RNA molecules. Most of the current RNA structure alignment programs are built on size-dependent scales, which complicate the interpretation of structure and functional relations. Meanwhile, the low speed prevents the programs from being applied to large-scale RNA structural database search. RESULTS: We developed an open-source algorithm, RNA-align, for RNA 3D structure alignment which has the structure similarity scaled by a size-independent and statistically interpretable scoring metric. Large-scale benchmark tests show that RNA-align significantly outperforms other state-of-the-art programs in both alignment accuracy and running speed. The major advantage of RNA-align lies at the quick convergence of the heuristic alignment iterations and the coarse-grained secondary structure assignment, both of which are crucial to the speed and accuracy of RNA structure alignments. AVAILABILITY AND IMPLEMENTATION: https://zhanglab.ccmb.med.umich.edu/RNA-align/. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

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.

MOTIVATION: Accurate prediction and interpretation of ligand bioactivities are essential for virtual screening and drug discovery. Unfortunately, many important drug targets lack experimental data about the ligand bioactivities; this is particularly true for G protein-coupled receptors (GPCRs), which account for the targets of about a third of drugs currently on the market. Computational approaches with the potential of precise assessment of ligand bioactivities and determination of key substructural features which determine ligand bioactivities are needed to address this issue. RESULTS: A new method, SED, was proposed to predict ligand bioactivities and to recognize key substructures associated with GPCRs through the coupling of screening for Lasso of long extended-connectivity fingerprints (ECFPs) with deep neural network training. The SED pipeline contains three successive steps: (i) representation of long ECFPs for ligand molecules, (ii) feature selection by screening for Lasso of ECFPs and (iii) bioactivity prediction through a deep neural network regression model. The method was examined on a set of 16 representative GPCRs that cover most subfamilies of human GPCRs, where each has 300-5000 ligand associations. The results show that SED achieves excellent performance in modelling ligand bioactivities, especially for those in the GPCR datasets without sufficient ligand associations, where SED improved the baseline predictors by 12% in correlation coefficient (r2) and 19% in root mean square error. Detail data analyses suggest that the major advantage of SED lies on its ability to detect substructures from long ECFPs which significantly improves the predictive performance. AVAILABILITY AND IMPLEMENTATION: The source code and datasets of SED are freely available at https://zhanglab.ccmb.med.umich.edu/SED/. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Accurate prediction of atomic-level protein structure is important for annotating the biological functions of protein molecules and for designing new compounds to regulate the functions. Template-based modeling (TBM), which aims to construct structural models by copying and refining the structural frameworks of other known proteins, remains the most accurate method for protein structure prediction. Due to the difficulty in recognizing distant-homology templates, however, the accuracy of TBM decreases rapidly when the evolutionary relationship between the query and template vanishes. In this study, we propose a new method, CEthreader, which first predicts residue-residue contacts by coupling evolutionary precision matrices with deep residual convolutional neural-networks. The predicted contact maps are then integrated with sequence profile alignments to recognize structural templates from the PDB. The method was tested on two independent benchmark sets consisting collectively of 1,153 non-homologous protein targets, where CEthreader detected 176% or 36% more correct templates with a TM-score >0.5 than the best state-of-the-art profile- or contact-based threading methods, respectively, for the Hard targets that lacked homologous templates. Moreover, CEthreader was able to identify 114% or 20% more correct templates with the same Fold as the query, after excluding structures from the same SCOPe Superfamily, than the best profile- or contact-based threading methods. Detailed analyses show that the major advantage of CEthreader lies in the efficient coupling of contact maps with profile alignments, which helps recognize global fold of protein structures when the homologous relationship between the query and template is weak. These results demonstrate an efficient new strategy to combine ab initio contact map prediction with profile alignments to significantly improve the accuracy of template-based structure prediction, especially for distant-homology proteins.

Motivation: As protein structure is more conserved than sequence during evolution, multiple structure alignment can be more informative than multiple sequence alignment, especially for distantly related proteins. With the rapid increase of the number of protein structures in the Protein Data Bank, it becomes urgent to develop efficient algorithms for multiple structure alignment. Results: A new multiple structure alignment algorithm (mTM-align) was proposed, which is an extension of the highly efficient pairwise structure alignment program TM-align. The algorithm was benchmarked on four widely used datasets, HOMSTRAD, SABmark_sup, SABmark_twi and SISY-multiple, showing that mTM-align consistently outperforms other algorithms. In addition, the comparison with the manually curated alignments in the HOMSTRAD database shows that the automated alignments built by mTM-align are in general more accurate. Therefore, mTM-align may be used as a reliable complement to construct multiple structure alignments for real-world applications. Availability and implementation: http://yanglab.nankai.edu.cn/mTM-align. Contact: [email protected] or [email protected]. Supplementary information: Supplementary data are available at Bioinformatics online.

Motivation: Sequence-order independent structural comparison, also called structural alignment, of small ligand molecules is often needed for computer-aided virtual drug screening. Although many ligand structure alignment programs are proposed, most of them build the alignments based on rigid-body shape comparison which cannot provide atom-specific alignment information nor allow structural variation; both abilities are critical to efficient high-throughput virtual screening. Results: We propose a novel ligand comparison algorithm, LS-align, to generate fast and accurate atom-level structural alignments of ligand molecules, through an iterative heuristic search of the target function that combines inter-atom distance with mass and chemical bond comparisons. LS-align contains two modules of Rigid-LS-align and Flexi-LS-align, designed for rigid-body and flexible alignments, respectively, where a ligand-size independent, statistics-based scoring function is developed to evaluate the similarity of ligand molecules relative to random ligand pairs. Large-scale benchmark tests are performed on prioritizing chemical ligands of 102 protein targets involving 1 415 871 candidate compounds from the DUD-E (Database of Useful Decoys: Enhanced) database, where LS-align achieves an average enrichment factor (EF) of 22.0 at the 1% cutoff and the AUC score of 0.75, which are significantly higher than other state-of-the-art methods. Detailed data analyses show that the advanced performance is mainly attributed to the design of the target function that combines structural and chemical information to enhance the sensitivity of recognizing subtle difference of ligand molecules and the introduces of structural flexibility that help capture the conformational changes induced by the ligand-receptor binding interactions. These data demonstrate a new avenue to improve the virtual screening efficiency through the development of sensitive ligand structural alignments. Availability and implementation: http://zhanglab.ccmb.med.umich.edu/LS-align/. Supplementary information: Supplementary data are available at Bioinformatics online.

Motivation: Precise assessment of ligand bioactivities (including IC50, EC50, Ki, Kd, etc.) is essential for virtual screening and lead compound identification. However, not all ligands have experimentally determined activities. In particular, many G protein-coupled receptors (GPCRs), which are the largest integral membrane protein family and represent targets of nearly 40% drugs on the market, lack published experimental data about ligand interactions. Computational methods with the ability to accurately predict the bioactivity of ligands can help efficiently address this problem. Results: We proposed a new method, WDL-RF, using weighted deep learning and random forest, to model the bioactivity of GPCR-associated ligand molecules. The pipeline of our algorithm consists of two consecutive stages: (i) molecular fingerprint generation through a new weighted deep learning method, and (ii) bioactivity calculations with a random forest model; where one uniqueness of the approach is that the model allows end-to-end learning of prediction pipelines with input ligands being of arbitrary size. The method was tested on a set of twenty-six non-redundant GPCRs that have a high number of active ligands, each with 200-4000 ligand associations. The results from our benchmark show that WDL-RF can generate bioactivity predictions with an average root-mean square error 1.33 and correlation coefficient (r2) 0.80 compared to the experimental measurements, which are significantly more accurate than the control predictors with different molecular fingerprints and descriptors. In particular, data-driven molecular fingerprint features, as extracted from the weighted deep learning models, can help solve deficiencies stemming from the use of traditional hand-crafted features and significantly increase the efficiency of short molecular fingerprints in virtual screening. Availability and implementation: The WDL-RF web server, as well as source codes and datasets of WDL-RF, is freely available at https://zhanglab.ccmb.med.umich.edu/WDL-RF/ for academic purposes. Supplementary information: Supplementary data are available at Bioinformatics online.

Accurate identification of the cancer types is essential to cancer diagnoses and treatments. Since cancer tissue and normal tissue have different gene expression, gene expression data can be used as an efficient feature source for cancer classification. However, accurate cancer classification directly using original gene expression profiles remains challenging due to the intrinsic high-dimension feature and the small size of the data samples. We proposed a new self-training subspace clustering algorithm under low-rank representation, called SSC-LRR, for cancer classification on gene expression data. Low-rank representation (LRR) is first applied to extract discriminative features from the high-dimensional gene expression data; the self-training subspace clustering (SSC) method is then used to generate the cancer classification predictions. The SSC-LRR was tested on two separate benchmark datasets in control with four state-of-the-art classification methods. It generated cancer classification predictions with an overall accuracy 89.7 percent and a general correlation 0.920, which are 18.9 and 24.4 percent higher than that of the best control method respectively. In addition, several genes (RNF114, HLA-DRB5, USP9Y, and PTPN20) were identified by SSC-LRR as new cancer identifiers that deserve further clinical investigation. Overall, the study demonstrated a new sensitive avenue to recognize cancer classifications from large-scale gene expression data.

MOTIVATION: Recent CASP experiments have witnessed exciting progress on folding large-size non-humongous proteins with the assistance of co-evolution based contact predictions. The success is however anecdotal due to the requirement of the contact prediction methods for the high volume of sequence homologs that are not available to most of the non-humongous protein targets. Development of efficient methods that can generate balanced and reliable contact maps for different type of protein targets is essential to enhance the success rate of the ab initio protein structure prediction. RESULTS: We developed a new pipeline, NeBcon, which uses the naïve Bayes classifier (NBC) theorem to combine eight state of the art contact methods that are built from co-evolution and machine learning approaches. The posterior probabilities of the NBC model are then trained with intrinsic structural features through neural network learning for the final contact map prediction. NeBcon was tested on 98 non-redundant proteins, which improves the accuracy of the best co-evolution based meta-server predictor by 22%; the magnitude of the improvement increases to 45% for the hard targets that lack sequence and structural homologs in the databases. Detailed data analysis showed that the major contribution to the improvement is due to the optimized NBC combination of the complementary information from both co-evolution and machine learning predictions. The neural network training also helps to improve the coupling of the NBC posterior probability and the intrinsic structural features, which were found particularly important for the proteins that do not have sufficient number of homologous sequences to derive reliable co-evolution profiles. AVAILIABLITY AND IMPLEMENTATION: On-line server and standalone package of the program are available at http://zhanglab.ccmb.med.umich.edu/NeBcon/ . CONTACT: [email protected]. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: Computationally generated non-native protein structure conformations (or decoys) are often used for designing protein folding simulation methods and force fields. However, almost all the decoy sets currently used in literature suffer from uneven root mean square deviation (RMSD) distribution with bias to non-protein like hydrogen-bonding and compactness patterns. Meanwhile, most protein decoy sets are pre-calculated and there is a lack of methods for automated generation of high-quality decoys for any target proteins. RESULTS: We developed a new algorithm, 3DRobot, to create protein structure decoys by free fragment assembly with enhanced hydrogen-bonding and compactness interactions. The method was benchmarked with three widely used decoy sets from ab initio folding and comparative modeling simulations. The decoys generated by 3DRobot are shown to have significantly enhanced diversity and evenness with a continuous distribution in the RMSD space. The new energy terms introduced in 3DRobot improve the hydrogen-bonding network and compactness of decoys, which eliminates the possibility of native structure recognition by trivial potentials. Algorithms that can automatically create such diverse and well-packed non-native conformations from any protein structure should have a broad impact on the development of advanced protein force field and folding simulation methods. AVAILIABLITY AND IMPLEMENTATION: http://zhanglab.ccmb.med.umich.edu/3DRobot/ CONTACT: [email protected]; [email protected] SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: More than half of proteins require binding of metal and acid radical ions for their structure and function. Identification of the ion-binding locations is important for understanding the biological functions of proteins. Due to the small size and high versatility of the metal and acid radical ions, however, computational prediction of their binding sites remains difficult. RESULTS: ) that are most frequently seen in protein databases. A sequence-based ab initio model is first trained on sequence profiles, where a modified AdaBoost algorithm is extended to balance binding and non-binding residue samples. A composite method IonCom is then developed to combine the ab initio model with multiple threading alignments for further improving the robustness of the binding site predictions. The pipeline was tested using 5-fold cross validations on a comprehensive set of 2,100 non-redundant proteins bound with 3,075 small ion ligands. Significant advantage was demonstrated compared with the state of the art ligand-binding methods including COACH and TargetS for high-accuracy ion-binding site identification. Detailed data analyses show that the major advantage of IonCom lies at the integration of complementary ab initio and template-based components. Ion-specific feature design and binding library selection also contribute to the improvement of small ion ligand binding predictions. AVAILABILITY AND IMPLEMENTATION: http://zhanglab.ccmb.med.umich.edu/IonCom CONTACT: [email protected] or [email protected] information: Supplementary data are available at Bioinformatics online.

MOTIVATION: Mutations in human genome are mainly through single nucleotide polymorphism, some of which can affect stability and function of proteins, causing human diseases. Several methods have been proposed to predict the effect of mutations on protein stability; but most require features from experimental structure. Given the fast progress in protein structure prediction, this work explores the possibility to improve the mutation-induced stability change prediction using low-resolution structure modeling. RESULTS: We developed a new method (STRUM) for predicting stability change caused by single-point mutations. Starting from wild-type sequences, 3D models are constructed by the iterative threading assembly refinement (I-TASSER) simulations, where physics- and knowledge-based energy functions are derived on the I-TASSER models and used to train STRUM models through gradient boosting regression. STRUM was assessed by 5-fold cross validation on 3421 experimentally determined mutations from 150 proteins. The Pearson correlation coefficient (PCC) between predicted and measured changes of Gibbs free-energy gap, ΔΔG, upon mutation reaches 0.79 with a root-mean-square error 1.2 kcal/mol in the mutation-based cross-validations. The PCC reduces if separating training and test mutations from non-homologous proteins, which reflects inherent correlations in the current mutation sample. Nevertheless, the results significantly outperform other state-of-the-art methods, including those built on experimental protein structures. Detailed analyses show that the most sensitive features in STRUM are the physics-based energy terms on I-TASSER models and the conservation scores from multiple-threading template alignments. However, the ΔΔG prediction accuracy has only a marginal dependence on the accuracy of protein structure models as long as the global fold is correct. These data demonstrate the feasibility to use low-resolution structure modeling for high-accuracy stability change prediction upon point mutations. AVAILABILITY AND IMPLEMENTATION: http://zhanglab.ccmb.med.umich.edu/STRUM/ CONTACT: [email protected] and [email protected] SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: G protein-coupled receptors (GPCRs) are probably the most attractive drug target membrane proteins, which constitute nearly half of drug targets in the contemporary drug discovery industry. While the majority of drug discovery studies employ existing GPCR and ligand interactions to identify new compounds, there remains a shortage of specific databases with precisely annotated GPCR-ligand associations. RESULTS: We have developed a new database, GLASS, which aims to provide a comprehensive, manually curated resource for experimentally validated GPCR-ligand associations. A new text-mining algorithm was proposed to collect GPCR-ligand interactions from the biomedical literature, which is then crosschecked with five primary pharmacological datasets, to enhance the coverage and accuracy of GPCR-ligand association data identifications. A special architecture has been designed to allow users for making homologous ligand search with flexible bioactivity parameters. The current database contains ∼500 000 unique entries, of which the vast majority stems from ligand associations with rhodopsin- and secretin-like receptors. The GLASS database should find its most useful application in various in silico GPCR screening and functional annotation studies. AVAILABILITY AND IMPLEMENTATION: The website of GLASS database is freely available at http://zhanglab.ccmb.med.umich.edu/GLASS/. CONTACT: [email protected] SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: There is a long-term interest in the challenging task of finding translocated and mislocated cancer biomarker proteins. Bioimages of subcellular protein distribution are new data sources which have attracted much attention in recent years because of their intuitive and detailed descriptions of protein distribution. However, automated methods in large-scale biomarker screening suffer significantly from the lack of subcellular location annotations for bioimages from cancer tissues. The transfer prediction idea of applying models trained on normal tissue proteins to predict the subcellular locations of cancerous ones is arbitrary because the protein distribution patterns may differ in normal and cancerous states. RESULTS: We developed a new semi-supervised protocol that can use unlabeled cancer protein data in model construction by an iterative and incremental training strategy. Our approach enables us to selectively use the low-quality images in normal states to expand the training sample space and provides a general way for dealing with the small size of annotated images used together with large unannotated ones. Experiments demonstrate that the new semi-supervised protocol can result in improved accuracy and sensitivity of subcellular location difference detection. AVAILABILITY AND IMPLEMENTATION: The data and code are available at: www.csbio.sjtu.edu.cn/bioinf/SemiBiomarker/. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: Cysteine-rich proteins cover many important families in nature but there are currently no methods specifically designed for modeling the structure of these proteins. The accuracy of disulfide connectivity pattern prediction, particularly for the proteins of higher-order connections, e.g., >3 bonds, is too low to effectively assist structure assembly simulations. RESULTS: We propose a new hierarchical order reduction protocol called Cyscon for disulfide-bonding prediction. The most confident disulfide bonds are first identified and bonding prediction is then focused on the remaining cysteine residues based on SVR training. Compared with purely machine learning-based approaches, Cyscon improved the average accuracy of connectivity pattern prediction by 21.9%. For proteins with more than 5 disulfide bonds, Cyscon improved the accuracy by 585% on the benchmark set of PDBCYS. When applied to 158 non-redundant cysteine-rich proteins, Cyscon predictions helped increase (or decrease) the TM-score (or RMSD) of the ab initio QUARK modeling by 12.1% (or 14.4%). This result demonstrates a new avenue to improve the ab initio structure modeling for cysteine-rich proteins. AVAILABILITY AND IMPLEMENTATION: http://www.csbio.sjtu.edu.cn/bioinf/Cyscon/ CONTACT: [email protected] or [email protected]. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

BACKGROUND: Structural comparison of protein-protein interfaces provides valuable insights into the functional relationship between proteins, which may not solely arise from shared evolutionary origin. A few methods that exist for such comparative studies have focused on structural models determined at atomic resolution, and may miss out interesting patterns present in large macromolecular complexes that are typically solved by low-resolution techniques. RESULTS: We developed a coarse-grained method, PCalign, to quantitatively evaluate physicochemical similarities between a given pair of protein-protein interfaces. This method uses an order-independent algorithm, geometric hashing, to superimpose the backbone atoms of a given pair of interfaces, and provides a normalized scoring function, PC-score, to account for the extent of overlap in terms of both geometric and chemical characteristics. We demonstrate that PCalign outperforms existing methods, and additionally facilitates comparative studies across models of different resolutions, which are not accommodated by existing methods. Furthermore, we illustrate potential application of our method to recognize interesting biological relationships masked by apparent lack of structural similarity. CONCLUSIONS: PCalign is a useful method in recognizing shared chemical and spatial patterns among protein-protein interfaces. It outperforms existing methods for high-quality data, and additionally facilitates comparison across structural models with different levels of details with proven robustness against noise.

The formation of protein-protein complexes is essential for proteins to perform their physiological functions in the cell. Mutations that prevent the proper formation of the correct complexes can have serious consequences for the associated cellular processes. Since experimental determination of protein-protein binding affinity remains difficult when performed on a large scale, computational methods for predicting the consequences of mutations on binding affinity are highly desirable. We show that a scoring function based on interface structure profiles collected from analogous protein-protein interactions in the PDB is a powerful predictor of protein binding affinity changes upon mutation. As a standalone feature, the differences between the interface profile score of the mutant and wild-type proteins has an accuracy equivalent to the best all-atom potentials, despite being two orders of magnitude faster once the profile has been constructed. Due to its unique sensitivity in collecting the evolutionary profiles of analogous binding interactions and the high speed of calculation, the interface profile score has additional advantages as a complementary feature to combine with physics-based potentials for improving the accuracy of composite scoring approaches. By incorporating the sequence-derived and residue-level coarse-grained potentials with the interface structure profile score, a composite model was constructed through the random forest training, which generates a Pearson correlation coefficient >0.8 between the predicted and observed binding free-energy changes upon mutation. This accuracy is comparable to, or outperforms in most cases, the current best methods, but does not require high-resolution full-atomic models of the mutant structures. The binding interface profiling approach should find useful application in human-disease mutation recognition and protein interface design studies.

MOTIVATION: Human cells are organized into compartments of different biochemical cellular processes. Having proteins appear at the right time to the correct locations in the cellular compartments is required to conduct their functions in normal cells, whereas mislocalization of proteins can result in pathological diseases, including cancer. RESULTS: To reveal the cancer-related protein mislocalizations, we developed an image-based multi-label subcellular location predictor, iLocator, which covers seven cellular localizations. The iLocator incorporates both global and local image descriptors and generates predictions by using an ensemble multi-label classifier. The algorithm has the ability to treat both single- and multiple-location proteins. We first trained and tested iLocator on 3240 normal human tissue images that have known subcellular location information from the human protein atlas. The iLocator was then used to generate protein localization predictions for 3696 protein images from seven cancer tissues that have no location annotations in the human protein atlas. By comparing the output data from normal and cancer tissues, we detected eight potential cancer biomarker proteins that have significant localization differences with P-value < 0.01. AVAILABILITY: http://www.csbio.sjtu.edu.cn/bioinf/iLocator/

MOTIVATION: Protein domains are subunits that can fold and evolve independently. Identification of domain boundary locations is often the first step in protein folding and function annotations. Most of the current methods deduce domain boundaries by sequence-based analysis, which has low accuracy. There is no efficient method for predicting discontinuous domains that consist of segments from separated sequence regions. As template-based methods are most efficient for protein 3D structure modeling, combining multiple threading alignment information should increase the accuracy and reliability of computational domain predictions. RESULT: We developed a new protein domain predictor, ThreaDom, which deduces domain boundary locations based on multiple threading alignments. The core of the method development is the derivation of a domain conservation score that combines information from template domain structures and terminal and internal alignment gaps. Tested on 630 non-redundant sequences, without using homologous templates, ThreaDom generates correct single- and multi-domain classifications in 81% of cases, where 78% have the domain linker assigned within ±20 residues. In a second test on 486 proteins with discontinuous domains, ThreaDom achieves an average precision 84% and recall 65% in domain boundary prediction. Finally, ThreaDom was examined on 56 targets from CASP8 and had a domain overlap rate 73, 87 and 85% with the target for Free Modeling, Hard multiple-domain and discontinuous domain proteins, respectively, which are significantly higher than most domain predictors in the CASP8. Similar results were achieved on the targets from the most recently CASP9 and CASP10 experiments. AVAILABILITY: http://zhanglab.ccmb.med.umich.edu/ThreaDom/. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: Residue-residue contacts across the transmembrane helices dictate the three-dimensional topology of alpha-helical membrane proteins. However, contact determination through experiments is difficult because most transmembrane proteins are hard to crystallize. RESULTS: We present a novel method (MemBrain) to derive transmembrane inter-helix contacts from amino acid sequences by combining correlated mutations and multiple machine learning classifiers. Tested on 60 non-redundant polytopic proteins using a strict leave-one-out cross-validation protocol, MemBrain achieves an average accuracy of 62%, which is 12.5% higher than the current best method from the literature. When applied to 13 recently solved G protein-coupled receptors, the MemBrain contact predictions helped increase the TM-score of the I-TASSER models by 37% in the transmembrane region. The number of foldable cases (TM-score >0.5) increased by 100%, where all G protein-coupled receptor templates and homologous templates with sequence identity >30% were excluded. These results demonstrate significant progress in contact prediction and a potential for contact-driven structure modeling of transmembrane proteins. AVAILABILITY: www.csbio.sjtu.edu.cn/bioinf/MemBrain/

MOTIVATION: Identification of protein-ligand binding sites is critical to protein function annotation and drug discovery. However, there is no method that could generate optimal binding site prediction for different protein types. Combination of complementary predictions is probably the most reliable solution to the problem. RESULTS: We develop two new methods, one based on binding-specific substructure comparison (TM-SITE) and another on sequence profile alignment (S-SITE), for complementary binding site predictions. The methods are tested on a set of 500 non-redundant proteins harboring 814 natural, drug-like and metal ion molecules. Starting from low-resolution protein structure predictions, the methods successfully recognize >51% of binding residues with average Matthews correlation coefficient (MCC) significantly higher (with P-value <10(-9) in student t-test) than other state-of-the-art methods, including COFACTOR, FINDSITE and ConCavity. When combining TM-SITE and S-SITE with other structure-based programs, a consensus approach (COACH) can increase MCC by 15% over the best individual predictions. COACH was examined in the recent community-wide COMEO experiment and consistently ranked as the best method in last 22 individual datasets with the Area Under the Curve score 22.5% higher than the second best method. These data demonstrate a new robust approach to protein-ligand binding site recognition, which is ready for genome-wide structure-based function annotations. AVAILABILITY: http://zhanglab.ccmb.med.umich.edu/COACH/

Computational protein design is a reverse procedure of protein folding and structure prediction, where constructing structures from evolutionarily related proteins has been demonstrated to be the most reliable method for protein 3-dimensional structure prediction. Following this spirit, we developed a novel method to design new protein sequences based on evolutionarily related protein families. For a given target structure, a set of proteins having similar fold are identified from the PDB library by structural alignments. A structural profile is then constructed from the protein templates and used to guide the conformational search of amino acid sequence space, where physicochemical packing is accommodated by single-sequence based solvation, torsion angle, and secondary structure predictions. The method was tested on a computational folding experiment based on a large set of 87 protein structures covering different fold classes, which showed that the evolution-based design significantly enhances the foldability and biological functionality of the designed sequences compared to the traditional physics-based force field methods. Without using homologous proteins, the designed sequences can be folded with an average root-mean-square-deviation of 2.1 Å to the target. As a case study, the method is extended to redesign all 243 structurally resolved proteins in the pathogenic bacteria Mycobacterium tuberculosis, which is the second leading cause of death from infectious disease. On a smaller scale, five sequences were randomly selected from the design pool and subjected to experimental validation. The results showed that all the designed proteins are soluble with distinct secondary structure and three have well ordered tertiary structure, as demonstrated by circular dichroism and NMR spectroscopy. Together, these results demonstrate a new avenue in computational protein design that uses knowledge of evolutionary conservation from protein structural families to engineer new protein molecules of improved fold stability and biological functionality.

MOTIVATION: Protein structure similarity is often measured by root mean squared deviation, global distance test score and template modeling score (TM-score). However, the scores themselves cannot provide information on how significant the structural similarity is. Also, it lacks a quantitative relation between the scores and conventional fold classifications. This article aims to answer two questions: (i) what is the statistical significance of TM-score? (ii) What is the probability of two proteins having the same fold given a specific TM-score? RESULTS: We first made an all-to-all gapless structural match on 6684 non-homologous single-domain proteins in the PDB and found that the TM-scores follow an extreme value distribution. The data allow us to assign each TM-score a P-value that measures the chance of two randomly selected proteins obtaining an equal or higher TM-score. With a TM-score at 0.5, for instance, its P-value is 5.5 x 10(-7), which means we need to consider at least 1.8 million random protein pairs to acquire a TM-score of no less than 0.5. Second, we examine the posterior probability of the same fold proteins from three datasets SCOP, CATH and the consensus of SCOP and CATH. It is found that the posterior probability from different datasets has a similar rapid phase transition around TM-score=0.5. This finding indicates that TM-score can be used as an approximate but quantitative criterion for protein topology classification, i.e. protein pairs with a TM-score >0.5 are mostly in the same fold while those with a TM-score <0.5 are mainly not in the same fold.

SUMMARY: G protein-coupled receptors (GPCRs) comprise the largest family of integral membrane proteins. They are the most important class of drug targets. While there exist crystal structures for only a very few GPCR sequences, numerous experiments have been performed on GPCRs to identify the critical residues and motifs. GPCRRD database is designed to systematically collect all experimental restraints (including residue orientation, contact and distance maps) available from the literature and primary GPCR resources using an automated text mining algorithm combined with manual validation, with the purpose of assisting GPCR 3D structure modeling and function annotation. The current dataset contains thousands of spatial restraints from mutagenesis, disulfide mapping distances, electron cryo-microscopy and Fourier-transform infrared spectroscopy experiments. AVAILABILITY: http://zhanglab.ccmb.med.umich.edu/GPCRRD/ CONTACT: [email protected] SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: Pair-wise residue-residue contacts in proteins can be predicted from both threading templates and sequence-based machine learning. However, most structure modeling approaches only use the template-based contact predictions in guiding the simulations; this is partly because the sequence-based contact predictions are usually considered to be less accurate than that by threading. With the rapid progress in sequence databases and machine-learning techniques, it is necessary to have a detailed and comprehensive assessment of the contact-prediction methods in different template conditions. RESULTS: We develop two methods for protein-contact predictions: SVM-SEQ is a sequence-based machine learning approach which trains a variety of sequence-derived features on contact maps; SVM-LOMETS collects consensus contact predictions from multiple threading templates. We test both methods on the same set of 554 proteins which are categorized into 'Easy', 'Medium', 'Hard' and 'Very Hard' targets based on the evolutionary and structural distance between templates and targets. For the Easy and Medium targets, SVM-LOMETS obviously outperforms SVM-SEQ; but for the Hard and Very Hard targets, the accuracy of the SVM-SEQ predictions is higher than that of SVM-LOMETS by 12-25%. If we combine the SVM-SEQ and SVM-LOMETS predictions together, the total number of correctly predicted contacts in the Hard proteins will increase by more than 60% (or 70% for the long-range contact with a sequence separation > or =24), compared with SVM-LOMETS alone. The advantage of SVM-SEQ is also shown in the CASP7 free modeling targets where the SVM-SEQ is around four times more accurate than SVM-LOMETS in the long-range contact prediction. These data demonstrate that the state-of-the-art sequence-based contact prediction has reached a level which may be helpful in assisting tertiary structure modeling for the targets which do not have close structure templates. The maximum yield should be obtained by the combination of both sequence- and template-based predictions.

BACKGROUND: Prediction of 3-dimensional protein structures from amino acid sequences represents one of the most important problems in computational structural biology. The community-wide Critical Assessment of Structure Prediction (CASP) experiments have been designed to obtain an objective assessment of the state-of-the-art of the field, where I-TASSER was ranked as the best method in the server section of the recent 7th CASP experiment. Our laboratory has since then received numerous requests about the public availability of the I-TASSER algorithm and the usage of the I-TASSER predictions. RESULTS: An on-line version of I-TASSER is developed at the KU Center for Bioinformatics which has generated protein structure predictions for thousands of modeling requests from more than 35 countries. A scoring function (C-score) based on the relative clustering structural density and the consensus significance score of multiple threading templates is introduced to estimate the accuracy of the I-TASSER predictions. A large-scale benchmark test demonstrates a strong correlation between the C-score and the TM-score (a structural similarity measurement with values in [0, 1]) of the first models with a correlation coefficient of 0.91. Using a C-score cutoff > -1.5 for the models of correct topology, both false positive and false negative rates are below 0.1. Combining C-score and protein length, the accuracy of the I-TASSER models can be predicted with an average error of 0.08 for TM-score and 2 A for RMSD. CONCLUSION: The I-TASSER server has been developed to generate automated full-length 3D protein structural predictions where the benchmarked scoring system helps users to obtain quantitative assessments of the I-TASSER models. The output of the I-TASSER server for each query includes up to five full-length models, the confidence score, the estimated TM-score and RMSD, and the standard deviation of the estimations. The I-TASSER server is freely available to the academic community at http://zhang.bioinformatics.ku.edu/I-TASSER.

G protein-coupled receptors (GPCRs), encoded by about 5% of human genes, comprise the largest family of integral membrane proteins and act as cell surface receptors responsible for the transduction of endogenous signal into a cellular response. Although tertiary structural information is crucial for function annotation and drug design, there are few experimentally determined GPCR structures. To address this issue, we employ the recently developed threading assembly refinement (TASSER) method to generate structure predictions for all 907 putative GPCRs in the human genome. Unlike traditional homology modeling approaches, TASSER modeling does not require solved homologous template structures; moreover, it often refines the structures closer to native. These features are essential for the comprehensive modeling of all human GPCRs when close homologous templates are absent. Based on a benchmarked confidence score, approximately 820 predicted models should have the correct folds. The majority of GPCR models share the characteristic seven-transmembrane helix topology, but 45 ORFs are predicted to have different structures. This is due to GPCR fragments that are predominantly from extracellular or intracellular domains as well as database annotation errors. Our preliminary validation includes the automated modeling of bovine rhodopsin, the only solved GPCR in the Protein Data Bank. With homologous templates excluded, the final model built by TASSER has a global C(alpha) root-mean-squared deviation from native of 4.6 angstroms, with a root-mean-squared deviation in the transmembrane helix region of 2.1 angstroms. Models of several representative GPCRs are compared with mutagenesis and affinity labeling data, and consistent agreement is demonstrated. Structure clustering of the predicted models shows that GPCRs with similar structures tend to belong to a similar functional class even when their sequences are diverse. These results demonstrate the usefulness and robustness of the in silico models for GPCR functional analysis. All predicted GPCR models are freely available for noncommercial users on our Web site (http://www.bioinformatics.buffalo.edu/GPCR).

MOTIVATION: Several protein function prediction methods employ structural features captured in three-dimensional (3D) descriptors of biologically relevant sites. These methods are successful when applied to high-resolution structures, but their detection ability in lower resolution predicted structures has only been tested for a few cases. RESULTS: A method that automatically generates a library of 3D functional descriptors for the structure-based prediction of enzyme active sites (automated functional templates, 593 in total for 162 different enzymes), based on functional and structural information automatically extracted from public databases, has been developed and evaluated using decoy structures. The applicability to predicted structures was investigated by analyzing decoys of varying quality, derived from enzyme native structures. For 35% of decoy structures, our method identifies the active site in models having 3-4 A coordinate root mean square deviation from the native structure, a quality that is reachable using state of the art protein structure prediction algorithms. AVAILABILITY: See http://www.bioinformatics.buffalo.edu/resources/aft/