Computational prediction of nucleic acid-binding residues in protein sequences is an active field of research, with over 80 methods that were released in the past 2 decades. We identify and discuss 87 sequence-based predictors that include dozens of recently published methods that are surveyed for the first time. We overview historical progress and examine multiple practical issues that include availability and impact of predictors, key features of their predictive models, and important aspects related to their training and assessment. We observe that the past decade has brought increased use of deep neural networks and protein language models, which contributed to substantial gains in the predictive performance. We also highlight advancements in vital and challenging issues that include cross-predictions between deoxyribonucleic acid (DNA)-binding and ribonucleic acid (RNA)-binding residues and targeting the two distinct sources of binding annotations, structure-based versus intrinsic disorder-based. The methods trained on the structure-annotated interactions tend to perform poorly on the disorder-annotated binding and vice versa, with only a few methods that target and perform well across both annotation types. The cross-predictions are a significant problem, with some predictors of DNA-binding or RNA-binding residues indiscriminately predicting interactions with both nucleic acid types. Moreover, we show that methods with web servers are cited substantially more than tools without implementation or with no longer working implementations, motivating the development and long-term maintenance of the web servers. We close by discussing future research directions that aim to drive further progress in this area.

Proteases contribute to a broad spectrum of cellular functions. Given a relatively limited amount of experimental data, developing accurate sequence-based predictors of substrate cleavage sites facilitates a better understanding of protease functions and substrate specificity. While many protease-specific predictors of substrate cleavage sites were developed, these efforts are outpaced by the growth of the protease substrate cleavage data. In particular, since data for 100+ protease types are available and this number continues to grow, it becomes impractical to publish predictors for new protease types, and instead it might be better to provide a computational platform that helps users to quickly and efficiently build predictors that address their specific needs. To this end, we conceptualized, developed, tested and released a versatile bioinformatics platform, ProsperousPlus, that empowers users, even those with no programming or little bioinformatics background, to build fast and accurate predictors of substrate cleavage sites. ProsperousPlus facilitates the use of the rapidly accumulating substrate cleavage data to train, empirically assess and deploy predictive models for user-selected substrate types. Benchmarking tests on test datasets show that our platform produces predictors that on average exceed the predictive performance of current state-of-the-art approaches. ProsperousPlus is available as a webserver and a stand-alone software package at http://prosperousplus.unimelb-biotools.cloud.edu.au/.

One of key features of intrinsically disordered regions (IDRs) is facilitation of protein-protein and protein-nucleic acids interactions. These disordered binding regions include molecular recognition features (MoRFs), short linear motifs (SLiMs) and longer binding domains. Vast majority of current predictors of disordered binding regions target MoRFs, with a handful of methods that predict SLiMs and disordered protein-binding domains. A new and broader class of disordered binding regions, linear interacting peptides (LIPs), was introduced recently and applied in the MobiDB resource. LIPs are segments in protein sequences that undergo disorder-to-order transition upon binding to a protein or a nucleic acid, and they cover MoRFs, SLiMs and disordered protein-binding domains. Although current predictors of MoRFs and disordered protein-binding regions could be used to identify some LIPs, there are no dedicated sequence-based predictors of LIPs. To this end, we introduce CLIP, a new predictor of LIPs that utilizes robust logistic regression model to combine three complementary types of inputs: co-evolutionary information derived from multiple sequence alignments, physicochemical profiles and disorder predictions. Ablation analysis suggests that the co-evolutionary information is particularly useful for this prediction and that combining the three inputs provides substantial improvements when compared to using these inputs individually. Comparative empirical assessments using low-similarity test datasets reveal that CLIP secures area under receiver operating characteristic curve (AUC) of 0.8 and substantially improves over the results produced by the closest current tools that predict MoRFs and disordered protein-binding regions. The webserver of CLIP is freely available at http://biomine.cs.vcu.edu/servers/CLIP/ and the standalone code can be downloaded from http://yanglab.qd.sdu.edu.cn/download/CLIP/.

Human leukocyte antigen class I (HLA-I) molecules bind intracellular peptides produced by protein hydrolysis and present them to the T cells for immune recognition and response. Prediction of peptides that bind HLA-I molecules is very important in immunotherapy. A growing number of computational predictors have been developed in recent years. We survey a comprehensive collection of 27 tools focusing on their input and output data characteristics, key aspects of the underlying predictive models and their availability. Moreover, we evaluate predictive performance for eight representative predictors. We consider a wide spectrum of relevant aspects including allele-specific analysis, influence of negative to positive data ratios and runtime. We also curate high-quality benchmark datasets based on analysis of the consistency of the data labels. Results reveal that each considered method provides accurate results, which can be explained by our analysis that finds that their predictive models capture meaningful binding motifs. Although some methods are overall more accurate than others, we find that none of them is universally superior. We provide a comprehensive comparison of the convenience as well as the accuracy of the methods under specific prediction scenarios, such as for specific alleles, metrics of predictive performance and constraints on runtime. Our systematic and broad analysis provides informative clues to the users to identify the most suitable tools for a given prediction scenario and for the developers to design future methods.

Proteins with intrinsically disordered regions (IDRs) are common among eukaryotes. Many IDRs interact with nucleic acids and proteins. Annotation of these interactions is supported by computational predictors, but to date, only one tool that predicts interactions with nucleic acids was released, and recent assessments demonstrate that current predictors offer modest levels of accuracy. We have developed DeepDISOBind, an innovative deep multi-task architecture that accurately predicts deoxyribonucleic acid (DNA)-, ribonucleic acid (RNA)- and protein-binding IDRs from protein sequences. DeepDISOBind relies on an information-rich sequence profile that is processed by an innovative multi-task deep neural network, where subsequent layers are gradually specialized to predict interactions with specific partner types. The common input layer links to a layer that differentiates protein- and nucleic acid-binding, which further links to layers that discriminate between DNA and RNA interactions. Empirical tests show that this multi-task design provides statistically significant gains in predictive quality across the three partner types when compared to a single-task design and a representative selection of the existing methods that cover both disorder- and structure-trained tools. Analysis of the predictions on the human proteome reveals that DeepDISOBind predictions can be encoded into protein-level propensities that accurately predict DNA- and RNA-binding proteins and protein hubs. DeepDISOBind is available at https://www.csuligroup.com/DeepDISOBind/.

Efforts to elucidate protein-DNA interactions at the molecular level rely in part on accurate predictions of DNA-binding residues in protein sequences. While there are over a dozen computational predictors of the DNA-binding residues, they are DNA-type agnostic and significantly cross-predict residues that interact with other ligands as DNA binding. We leverage a custom-designed machine learning architecture to introduce DNAgenie, first-of-its-kind predictor of residues that interact with A-DNA, B-DNA and single-stranded DNA. DNAgenie uses a comprehensive physiochemical profile extracted from an input protein sequence and implements a two-step refinement process to provide accurate predictions and to minimize the cross-predictions. Comparative tests on an independent test dataset demonstrate that DNAgenie outperforms the current methods that we adapt to predict residue-level interactions with the three DNA types. Further analysis finds that the use of the second (refinement) step leads to a substantial reduction in the cross predictions. Empirical tests show that DNAgenie's outputs that are converted to coarse-grained protein-level predictions compare favorably against recent tools that predict which DNA-binding proteins interact with double-stranded versus single-stranded DNAs. Moreover, predictions from the sequences of the whole human proteome reveal that the results produced by DNAgenie substantially overlap with the known DNA-binding proteins while also including promising leads for several hundred previously unknown putative DNA binders. These results suggest that DNAgenie is a valuable tool for the sequence-based characterization of protein functions. The DNAgenie's webserver is available at http://biomine.cs.vcu.edu/servers/DNAgenie/.

In recent years, high-throughput experimental techniques have significantly enhanced the accuracy and coverage of protein-protein interaction identification, including human-pathogen protein-protein interactions (HP-PPIs). Despite this progress, experimental methods are, in general, expensive in terms of both time and labour costs, especially considering that there are enormous amounts of potential protein-interacting partners. Developing computational methods to predict interactions between human and bacteria pathogen has thus become critical and meaningful, in both facilitating the detection of interactions and mining incomplete interaction maps. In this paper, we present a systematic evaluation of machine learning-based computational methods for human-bacterium protein-protein interactions (HB-PPIs). We first reviewed a vast number of publicly available databases of HP-PPIs and then critically evaluate the availability of these databases. Benefitting from its well-structured nature, we subsequently preprocess the data and identified six bacterium pathogens that could be used to study bacterium subjects in which a human was the host. Additionally, we thoroughly reviewed the literature on 'host-pathogen interactions' whereby existing models were summarized that we used to jointly study the impact of different feature representation algorithms and evaluate the performance of existing machine learning computational models. Owing to the abundance of sequence information and the limited scale of other protein-related information, we adopted the primary protocol from the literature and dedicated our analysis to a comprehensive assessment of sequence information and machine learning models. A systematic evaluation of machine learning models and a wide range of feature representation algorithms based on sequence information are presented as a comparison survey towards the prediction performance evaluation of HB-PPIs.

MOTIVATION: X-ray crystallography was used to produce nearly 90% of protein structures. These efforts were supported by numerous sequence-based tools that accurately predict crystallizable proteins. However, protein structures vary widely in their quality, typically measured with resolution and R-free. This impacts the ability to use these structures for some applications including rational drug design and molecular docking and motivates development of methods that accurately predict structure quality from sequence. RESULTS: We introduce XRRpred, the first predictor of the resolution and R-free values from protein sequences. XRRpred relies on original sequence profiles, hand-crafted features, empirically selected and parametrized regressors and modern resampling techniques. Using an independent test dataset, we show that XRRpred provides accurate predictions of resolution and R-free. We demonstrate that XRRpred's predictions correctly model relationship between the resolution and R-free and reproduce structure quality relations between structural classes of proteins. We also show that XRRpred significantly outperforms indirect alternative ways to predict the structure quality that include predictors of crystallization propensity and an alignment-based approach. XRRpred is available as a convenient webserver that allows batch predictions and offers informative visualization of the results. AVAILABILITY AND IMPLEMENTATION: http://biomine.cs.vcu.edu/servers/XRRPred/. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: Intrinsically disordered protein regions interact with proteins, nucleic acids and lipids. Regions that bind lipids are implicated in a wide spectrum of cellular functions and several human diseases. Motivated by the growing amount of experimental data for these interactions and lack of tools that can predict them from the protein sequence, we develop DisoLipPred, the first predictor of the disordered lipid-binding residues (DLBRs). RESULTS: DisoLipPred relies on a deep bidirectional recurrent network that implements three innovative features: transfer learning, bypass module that sidesteps predictions for putative structured residues, and expanded inputs that cover physiochemical properties associated with the protein-lipid interactions. Ablation analysis shows that these features drive predictive quality of DisoLipPred. Tests on an independent test dataset and the yeast proteome reveal that DisoLipPred generates accurate results and that none of the related existing tools can be used to indirectly identify DLBR. We also show that DisoLipPred's predictions complement the results generated by predictors of the transmembrane regions. Altogether, we conclude that DisoLipPred provides high-quality predictions of DLBRs that complement the currently available methods. AVAILABILITY AND IMPLEMENTATION: DisoLipPred's webserver is available at http://biomine.cs.vcu.edu/servers/DisoLipPred/. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Experimental annotations of intrinsic disorder are available for 0.1% of 147 000 000 of currently sequenced proteins. Over 60 sequence-based disorder predictors were developed to help bridge this gap. Current benchmarks of these methods assess predictive performance on datasets of proteins; however, predictions are often interpreted for individual proteins. We demonstrate that the protein-level predictive performance varies substantially from the dataset-level benchmarks. Thus, we perform first-of-its-kind protein-level assessment for 13 popular disorder predictors using 6200 disorder-annotated proteins. We show that the protein-level distributions are substantially skewed toward high predictive quality while having long tails of poor predictions. Consequently, between 57% and 75% proteins secure higher predictive performance than the currently used dataset-level assessment suggests, but as many as 30% of proteins that are located in the long tails suffer low predictive performance. These proteins typically have relatively high amounts of disorder, in contrast to the mostly structured proteins that are predicted accurately by all 13 methods. Interestingly, each predictor provides the most accurate results for some number of proteins, while the best-performing at the dataset-level method is in fact the best for only about 30% of proteins. Moreover, the majority of proteins are predicted more accurately than the dataset-level performance of the most accurate tool by at least four disorder predictors. While these results suggests that disorder predictors outperform their current benchmark performance for the majority of proteins and that they complement each other, novel tools that accurately identify the hard-to-predict proteins and that make accurate predictions for these proteins are needed.

Briefings in Bioinformatics, 2017. https://doi.org/10.1093/bib/bbx168 In the original version of this article, reference 60 incorrectly listed Ben-Tal N as the third author of the paper ‘Local geometry and evolutionary conservation of protein surfaces reveal the multiple recognition patches in protein-protein interactions’. This reference has now been corrected to the below: Laine E, Carbone A. Local geometry and evolutionary conservation of protein surfaces reveal the multiple recognition patches in protein-protein interactions. PLoS Comput Biol 2015;11(12):e1004580. The authors would like to apologize for this error.

MOTIVATION: Proteases are enzymes that cleave target substrate proteins by catalyzing the hydrolysis of peptide bonds between specific amino acids. While the functional proteolysis regulated by proteases plays a central role in the 'life and death' cellular processes, many of the corresponding substrates and their cleavage sites were not found yet. Availability of accurate predictors of the substrates and cleavage sites would facilitate understanding of proteases' functions and physiological roles. Deep learning is a promising approach for the development of accurate predictors of substrate cleavage events. RESULTS: We propose DeepCleave, the first deep learning-based predictor of protease-specific substrates and cleavage sites. DeepCleave uses protein substrate sequence data as input and employs convolutional neural networks with transfer learning to train accurate predictive models. High predictive performance of our models stems from the use of high-quality cleavage site features extracted from the substrate sequences through the deep learning process, and the application of transfer learning, multiple kernels and attention layer in the design of the deep network. Empirical tests against several related state-of-the-art methods demonstrate that DeepCleave outperforms these methods in predicting caspase and matrix metalloprotease substrate-cleavage sites. AVAILABILITY AND IMPLEMENTATION: The DeepCleave webserver and source code are freely available at http://deepcleave.erc.monash.edu/. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: There are over 30 sequence-based predictors of the protein-binding residues (PBRs). They use either structure-annotated or disorder-annotated training datasets, potentially creating a dichotomy where the structure-/disorder-specific models may not be able to cross-over to accurately predict the other type. Moreover, the structure-trained predictors were shown to substantially cross-predict PBRs among residues that interact with non-protein partners (nucleic acids and small ligands). We address these issues by performing first-of-its-kind comparative study of a representative collection of disorder- and structure-trained predictors using a comprehensive benchmark set with the structure- and disorder-derived annotations of PBRs (to analyze the cross-over) and the protein-, nucleic acid- and small ligand-binding proteins (to study the cross-predictions). RESULTS: Three predictors provide accurate results: SCRIBER, ANCHOR and disoRDPbind. Some of the structure-trained methods make accurate predictions on the structure-annotated proteins. Similarly, the disorder-trained predictors predict well on the disorder-annotated proteins. However, the considered predictors generally fail to cross-over, with the exception of SCRIBER. Our study also reveals that virtually all methods substantially cross-predict PBRs, except for SCRIBER for the structure-annotated proteins and disoRDPbind for the disorder-annotated proteins. We formulate a novel hybrid predictor, hybridPBRpred, that combines results produced by disoRDPbind and SCRIBER to accurately predict disorder- and structure-annotated PBRs. HybridPBRpred generates accurate results that cross-over structure- and disorder-annotated proteins and produces relatively low amount of cross-predictions, offering an accurate alternative to predict PBRs. AVAILABILITY AND IMPLEMENTATION: HybridPBRpred webserver, benchmark dataset and supplementary information are available at http://biomine.cs.vcu.edu/servers/hybridPBRpred/. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: Knowledge of protein-binding residues (PBRs) improves our understanding of protein-protein interactions, contributes to the prediction of protein functions and facilitates protein-protein docking calculations. While many sequence-based predictors of PBRs were published, they offer modest levels of predictive performance and most of them cross-predict residues that interact with other partners. One unexplored option to improve the predictive quality is to design consensus predictors that combine results produced by multiple methods. RESULTS: We empirically investigate predictive performance of a representative set of nine predictors of PBRs. We report substantial differences in predictive quality when these methods are used to predict individual proteins, which contrast with the dataset-level benchmarks that are currently used to assess and compare these methods. Our analysis provides new insights for the cross-prediction concern, dissects complementarity between predictors and demonstrates that predictive performance of the top methods depends on unique characteristics of the input protein sequence. Using these insights, we developed PROBselect, first-of-its-kind consensus predictor of PBRs. Our design is based on the dynamic predictor selection at the protein level, where the selection relies on regression-based models that accurately estimate predictive performance of selected predictors directly from the sequence. Empirical assessment using a low-similarity test dataset shows that PROBselect provides significantly improved predictive quality when compared with the current predictors and conventional consensuses that combine residue-level predictions. Moreover, PROBselect informs the users about the expected predictive quality for the prediction generated from a given input protein. AVAILABILITY AND IMPLEMENTATION: PROBselect is available at http://bioinformatics.csu.edu.cn/PROBselect/home/index. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

While protein-DNA interactions are crucial for a wide range of cellular functions, only a small fraction of these interactions was annotated to date. One solution to close this annotation gap is to employ computational methods that accurately predict protein-DNA interactions from widely available protein sequences. We present and empirically test first-of-its-kind predictor of DNA-binding residues in local segments of protein sequences that relies on the Fuzzy Cognitive Map (FCM) model. The FCM model uses information about putative solvent accessibility, evolutionary conservation, and relative propensities of amino acid to interact with DNA to generate putative DNA-binding residues. Empirical tests on a benchmark dataset reveal that the FCM model secures AUC = 0.72 and outperforms recently released hybridNAP predictor and several popular machine learning methods including Support Vector Machines, Naïve Bayes, and k-Nearest Neighbor. The improvements in the predictive performance result from an intrinsic feature of FCMs that incorporate relations between the input features, besides the relations between the inputs and output that are modelled by other algorithms. We also empirically demonstrate that use of a short sliding window results in further improvements in the predictive quality. The funDNApred webserver that implements the FCM predictor is available at http://biomine.cs.vcu.edu/servers/funDNApred/.

Drug-protein interactions (DPIs) underlie the desired therapeutic actions and the adverse side effects of a significant majority of drugs. Computational prediction of DPIs facilitates research in drug discovery, characterization and repurposing. Similarity-based methods that do not require knowledge of protein structures are particularly suitable for druggable genome-wide predictions of DPIs. We review 35 high-impact similarity-based predictors that were published in the past decade. We group them based on three types of similarities and their combinations that they use. We discuss and compare key aspects of these methods including source databases, internal databases and their predictive models. Using our novel benchmark database, we perform comparative empirical analysis of predictive performance of seven types of representative predictors that utilize each type of similarity individually and all possible combinations of similarities. We assess predictive quality at the database-wide DPI level and we are the first to also include evaluation over individual drugs. Our comprehensive analysis shows that predictors that use more similarity types outperform methods that employ fewer similarities, and that the model combining all three types of similarities secures area under the receiver operating characteristic curve of 0.93. We offer a comprehensive analysis of sensitivity of predictive performance to intrinsic and extrinsic characteristics of the considered predictors. We find that predictive performance is sensitive to low levels of similarities between sequences of the drug targets and several extrinsic properties of the input drug structures, drug profiles and drug targets. The benchmark database and a webserver for the seven predictors are freely available at http://biomine.cs.vcu.edu/servers/CONNECTOR/.

Proteins interact with a variety of molecules including proteins and nucleic acids. We review a comprehensive collection of over 50 studies that analyze and/or predict these interactions. While majority of these studies address either solely protein-DNA or protein-RNA binding, only a few have a wider scope that covers both protein-protein and protein-nucleic acid binding. Our analysis reveals that binding residues are typically characterized with three hallmarks: relative solvent accessibility (RSA), evolutionary conservation and propensity of amino acids (AAs) for binding. Motivated by drawbacks of the prior studies, we perform a large-scale analysis to quantify and contrast the three hallmarks for residues that bind DNA-, RNA-, protein- and (for the first time) multi-ligand-binding residues that interact with DNA and proteins, and with RNA and proteins. Results generated on a well-annotated data set of over 23 000 proteins show that conservation of binding residues is higher for nucleic acid- than protein-binding residues. Multi-ligand-binding residues are more conserved and have higher RSA than single-ligand-binding residues. We empirically show that each hallmark discriminates between binding and nonbinding residues, even predicted RSA, and that combining them improves discriminatory power for each of the five types of interactions. Linear scoring functions that combine these hallmarks offer good predictive performance of residue-level propensity for binding and provide intuitive interpretation of predictions. Better understanding of these residue-level interactions will facilitate development of methods that accurately predict binding in the exponentially growing databases of protein sequences.

MOTIVATION: While putative intrinsic disorder is widely used, none of the predictors provides quality assessment (QA) scores. QA scores estimate the likelihood that predictions are correct at a residue level and have been applied in other bioinformatics areas. We recently reported that QA scores derived from putative disorder propensities perform relatively poorly for native disordered residues. Here we design and validate a general approach to construct QA predictors for disorder predictions. RESULTS: The QUARTER (QUality Assessment for pRotein inTrinsic disordEr pRedictions) toolbox of methods accommodates a diverse set of ten disorder predictors. It builds upon several innovative design elements including use and scaling of selected physicochemical properties of the input sequence, post-processing of disorder propensity scores, and a feature selection that optimizes the predictive models to a specific disorder predictor. We empirically establish that each one of these elements contributes to the overall predictive performance of our tool and that QUARTER's outputs significantly outperform QA scores derived from the outputs generated the disorder predictors. The best performing QA scores for a single disorder predictor identify 13% of residues that are predicted with 98% precision. QA scores computed by combining results of the ten disorder predictors cover 40% of residues with 95% precision. Case studies are used to show how to interpret the QA scores. QA scores based on the high precision combined predictions are applied to analyze disorder in the human proteome. AVAILABILITY AND IMPLEMENTATION: http://biomine.cs.vcu.edu/servers/QUARTER/. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: Accurate predictions of protein-binding residues (PBRs) enhances understanding of molecular-level rules governing protein-protein interactions, helps protein-protein docking and facilitates annotation of protein functions. Recent studies show that current sequence-based predictors of PBRs severely cross-predict residues that interact with other types of protein partners (e.g. RNA and DNA) as PBRs. Moreover, these methods are relatively slow, prohibiting genome-scale use. RESULTS: We propose a novel, accurate and fast sequence-based predictor of PBRs that minimizes the cross-predictions. Our SCRIBER (SeleCtive pRoteIn-Binding rEsidue pRedictor) method takes advantage of three innovations: comprehensive dataset that covers multiple types of binding residues, novel types of inputs that are relevant to the prediction of PBRs, and an architecture that is tailored to reduce the cross-predictions. The dataset includes complete protein chains and offers improved coverage of binding annotations that are transferred from multiple protein-protein complexes. We utilize innovative two-layer architecture where the first layer generates a prediction of protein-binding, RNA-binding, DNA-binding and small ligand-binding residues. The second layer re-predicts PBRs by reducing overlap between PBRs and the other types of binding residues produced in the first layer. Empirical tests on an independent test dataset reveal that SCRIBER significantly outperforms current predictors and that all three innovations contribute to its high predictive performance. SCRIBER reduces cross-predictions by between 41% and 69% and our conservative estimates show that it is at least 3 times faster. We provide putative PBRs produced by SCRIBER for the entire human proteome and use these results to hypothesize that about 14% of currently known human protein domains bind proteins. AVAILABILITY AND IMPLEMENTATION: SCRIBER webserver is available at http://biomine.cs.vcu.edu/servers/SCRIBER/. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

X-ray crystallography is the main tool for structural determination of proteins. Yet, the underlying crystallization process is costly, has a high attrition rate and involves a series of trial-and-error attempts to obtain diffraction-quality crystals. The Structural Genomics Consortium aims to systematically solve representative structures of major protein-fold classes using primarily high-throughput X-ray crystallography. The attrition rate of these efforts can be improved by selection of proteins that are potentially easier to be crystallized. In this context, bioinformatics approaches have been developed to predict crystallization propensities based on protein sequences. These approaches are used to facilitate prioritization of the most promising target proteins, search for alternative structural orthologues of the target proteins and suggest designs of constructs capable of potentially enhancing the likelihood of successful crystallization. We reviewed and compared nine predictors of protein crystallization propensity. Moreover, we demonstrated that integrating selected outputs from multiple predictors as candidate input features to build the predictive model results in a significantly higher predictive performance when compared to using these predictors individually. Furthermore, we also introduced a new and accurate predictor of protein crystallization propensity, Crysf, which uses functional features extracted from UniProt as inputs. This comprehensive review will assist structural biologists in selecting the most appropriate predictor, and is also beneficial for bioinformaticians to develop a new generation of predictive algorithms.

Understanding of molecular mechanisms that govern protein-protein interactions and accurate modeling of protein-protein docking rely on accurate identification and prediction of protein-binding partners and protein-binding residues. We review over 40 methods that predict protein-protein interactions from protein sequences including methods that predict interacting protein pairs, protein-binding residues for a pair of interacting sequences and protein-binding residues in a single protein chain. We focus on the latter methods that provide residue-level annotations and that can be broadly applied to all protein sequences. We compare their architectures, inputs and outputs, and we discuss aspects related to their assessment and availability. We also perform first-of-its-kind comprehensive empirical comparison of representative predictors of protein-binding residues using a novel and high-quality benchmark data set. We show that the selected predictors accurately discriminate protein-binding and non-binding residues and that newer methods outperform older designs. However, these methods are unable to accurately separate residues that bind other molecules, such as DNA, RNA and small ligands, from the protein-binding residues. This cross-prediction, defined as the incorrect prediction of nucleic-acid- and small-ligand-binding residues as protein binding, is substantial for all evaluated methods and is not driven by the proximity to the native protein-binding residues. We discuss reasons for this drawback and we offer several recommendations. In particular, we postulate the need for a new generation of more accurate predictors and data sets, inclusion of a comprehensive assessment of the cross-predictions in future studies and higher standards of availability of the published methods.

Briefings in Bioinformatics 2017. doi: 10.1093/bib/bbx018. The authors would like to change the wording in their ‘Results and discussion’ section to remove the names of ‘ParCrys’ and ‘XANNpred’, which were identified as not having active webservers. The webservers for ParCrys and XANNpred are active and can be found at http://www.compbio.dundee.ac.uk/xtal and http://www.compbio.dundee.ac.uk/xannpred respectively. The authors apologize to the owners for the error which has been corrected.

BACKGROUND: Development of predictors of propensity of protein sequences for successful crystallization has been actively pursued for over a decade. A few novel methods that expanded the scope of these predictions to address additional steps of protein production and structure determination pipelines were released in recent years. The predictive performance of the current methods is modest. This is because the only input that they use is the protein sequence and since the experimental annotations of these data might be inconsistent given that they were collected across many laboratories and centers. However, even these modest levels of predictive quality are still practical compared to the reported low success rates of crystallization, which are below 10%. We focus on another important aspect related to a high computational cost of running the predictors that offer the expanded scope. RESULTS: We introduce a novel fDETECT webserver that provides very fast and modestly accurate predictions of the success of protein production, purification, crystallization, and structure determination. Empirical tests on two datasets demonstrate that fDETECT is more accurate than the only other similarly fast method, and similarly accurate and three orders of magnitude faster than the currently most accurate predictors. Our method predicts a single protein in about 120 milliseconds and needs less than an hour to generate the four predictions for an entire human proteome. Moreover, we empirically show that fDETECT secures similar levels of predictive performance when compared with four representative methods that only predict success of crystallization, while it also provides the other three predictions. A webserver that implements fDETECT is available at http://biomine.cs.vcu.edu/servers/fDETECT/ . CONCLUSIONS: fDETECT is a computational tool that supports target selection for protein production and X-ray crystallography-based structure determination. It offers predictive quality that matches or exceeds other state-of-the-art tools and is especially suitable for the analysis of large protein sets.

Motivated by the pressing need to characterize protein-DNA and protein-RNA interactions on large scale, we review a comprehensive set of 30 computational methods for high-throughput prediction of RNA- or DNA-binding residues from protein sequences. We summarize these predictors from several significant perspectives including their design, outputs and availability. We perform empirical assessment of methods that offer web servers using a new benchmark data set characterized by a more complete annotation that includes binding residues transferred from the same or similar proteins. We show that predictors of DNA-binding (RNA-binding) residues offer relatively strong predictive performance but they are unable to properly separate DNA- from RNA-binding residues. We design and empirically assess several types of consensuses and demonstrate that machine learning (ML)-based approaches provide improved predictive performance when compared with the individual predictors of DNA-binding residues or RNA-binding residues. We also formulate and execute first-of-its-kind study that targets combined prediction of DNA- and RNA-binding residues. We design and test three types of consensuses for this prediction and conclude that this novel approach that relies on ML design provides better predictive quality than individual predictors when tested on prediction of DNA- and RNA-binding residues individually. It also substantially improves discrimination between these two types of nucleic acids. Our results suggest that development of a new generation of predictors would benefit from using training data sets that combine both RNA- and DNA-binding proteins, designing new inputs that specifically target either DNA- or RNA-binding residues and pursuing combined prediction of DNA- and RNA-binding residues.

MOTIVATION: Disordered flexible linkers (DFLs) are disordered regions that serve as flexible linkers/spacers in multi-domain proteins or between structured constituents in domains. They are different from flexible linkers/residues because they are disordered and longer. Availability of experimentally annotated DFLs provides an opportunity to build high-throughput computational predictors of these regions from protein sequences. To date, there are no computational methods that directly predict DFLs and they can be found only indirectly by filtering predicted flexible residues with predictions of disorder. RESULTS: We conceptualized, developed and empirically assessed a first-of-its-kind sequence-based predictor of DFLs, DFLpred. This method outputs propensity to form DFLs for each residue in the input sequence. DFLpred uses a small set of empirically selected features that quantify propensities to form certain secondary structures, disordered regions and structured regions, which are processed by a fast linear model. Our high-throughput predictor can be used on the whole-proteome scale; it needs <1 h to predict entire proteome on a single CPU. When assessed on an independent test dataset with low sequence-identity proteins, it secures area under the receiver operating characteristic curve equal 0.715 and outperforms existing alternatives that include methods for the prediction of flexible linkers, flexible residues, intrinsically disordered residues and various combinations of these methods. Prediction on the complete human proteome reveals that about 10% of proteins have a large content of over 30% DFL residues. We also estimate that about 6000 DFL regions are long with ≥30 consecutive residues. AVAILABILITY AND IMPLEMENTATION: http://biomine.ece.ualberta.ca/DFLpred/ CONTACT: [email protected] SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: Many drugs interact with numerous proteins besides their intended therapeutic targets and a substantial portion of these interactions is yet to be elucidated. Protein-Drug Interaction Database (PDID) addresses incompleteness of these data by providing access to putative protein-drug interactions that cover the entire structural human proteome. RESULTS: PDID covers 9652 structures from 3746 proteins and houses 16 800 putative interactions generated from close to 1.1 million accurate, all-atom structure-based predictions for several dozens of popular drugs. The predictions were generated with three modern methods: ILbind, SMAP and eFindSite. They are accompanied by propensity scores that quantify likelihood of interactions and coordinates of the putative location of the binding drugs in the corresponding protein structures. PDID complements the current databases that focus on the curated interactions and the BioDrugScreen database that relies on docking to find putative interactions. Moreover, we also include experimentally curated interactions which are linked to their sources: DrugBank, BindingDB and Protein Data Bank. Our database can be used to facilitate studies related to polypharmacology of drugs including repurposing and explaining side effects of drugs. AVAILABILITY AND IMPLEMENTATION: PDID database is freely available at http://biomine.ece.ualberta.ca/PDID/.

MicroRNAs (miRNAs) are short endogenous noncoding RNAs that bind to target mRNAs, usually resulting in degradation and translational repression. Identification of miRNA targets is crucial for deciphering functional roles of the numerous miRNAs that are rapidly generated by sequencing efforts. Computational prediction methods are widely used for high-throughput generation of putative miRNA targets. We review a comprehensive collection of 38 miRNA sequence-based computational target predictors in animals that were developed over the past decade. Our in-depth analysis considers all significant perspectives including the underlying predictive methodologies with focus on how they draw from the mechanistic basis of the miRNA-mRNA interaction. We also discuss ease of use, availability, impact of the considered predictors and the evaluation protocols that were used to assess them. We are the first to comparatively and comprehensively evaluate seven representative methods when predicting miRNA targets at the duplex and gene levels. The gene-level evaluation is based on three benchmark data sets that rely on different ways to annotate targets including biochemical assays, microarrays and pSILAC. We offer practical advice on selection of appropriate predictors according to certain properties of miRNA sequences, characteristics of a specific application and desired levels of predictive quality. We also discuss future work related to the design of new models, data quality, improved usability, need for standardized evaluation and ability to predict mRNA expression changes.

MOTIVATION: Off-target interactions of a popular immunosuppressant Cyclosporine A (CSA) with several proteins besides its molecular target, cyclophilin A, are implicated in the activation of signaling pathways that lead to numerous side effects of this drug. RESULTS: Using structural human proteome and a novel algorithm for inverse ligand binding prediction, ILbind, we determined a comprehensive set of 100+ putative partners of CSA. We empirically show that predictive quality of ILbind is better compared with other available predictors for this compound. We linked the putative target proteins, which include many new partners of CSA, with cellular functions, canonical pathways and toxicities that are typical for patients who take this drug. We used complementary approaches (molecular docking, molecular dynamics, surface plasmon resonance binding analysis and enzymatic assays) to validate and characterize three novel CSA targets: calpain 2, caspase 3 and p38 MAP kinase 14. The three targets are involved in the apoptotic pathways, are interconnected and are implicated in nephrotoxicity.

MOTIVATION: Gaussian network model (GNM) is widely adopted to analyze and understand protein dynamics, function and conformational changes. The existing GNM-based approaches require atomic coordinates of the corresponding protein and cannot be used when only the sequence is known. RESULTS: We report, first of its kind, GNM model that allows modeling using the sequence. Our linear regression-based, parameter-free, sequence-derived GNM (L-pfSeqGNM) uses contact maps predicted from the sequence and models local, in the sequence, contact neighborhoods with the linear regression. Empirical benchmarking shows relatively high correlations between the native and the predicted with L-pfSeqGNM B-factors and between the cross-correlations of residue fluctuations derived from the structure- and the sequence-based GNM models. Our results demonstrate that L-pfSeqGNM is an attractive platform to explore protein dynamics. In contrast to the highly used GNMs that require protein structures that number in thousands, our model can be used to study motions for the millions of the readily available sequences, which finds applications in modeling conformational changes, protein-protein interactions and protein functions.

Multiple-instance learning (MIL) is a supervised learning technique that addresses the problem of classifying bags of instances instead of single instances. In this paper, we introduce a rule-based MIL algorithm, called mi-DS, and compare it with 21 existing MIL algorithms on 26 commonly used data sets. The results show that mi-DS performs on par with or better than several well-known algorithms and generates models characterized by balanced values of precision and recall. Importantly, the introduced method provides a framework that can be used for converting other rule-based algorithms into MIL algorithms.

MOTIVATION: Nucleotides are multifunctional molecules that are essential for numerous biological processes. They serve as sources for chemical energy, participate in the cellular signaling and they are involved in the enzymatic reactions. The knowledge of the nucleotide-protein interactions helps with annotation of protein functions and finds applications in drug design. RESULTS: We propose a novel ensemble of accurate high-throughput predictors of binding residues from the protein sequence for ATP, ADP, AMP, GTP and GDP. Empirical tests show that our NsitePred method significantly outperforms existing predictors and approaches based on sequence alignment and residue conservation scoring. The NsitePred accurately finds more binding residues and binding sites and it performs particularly well for the sites with residues that are clustered close together in the sequence. The high predictive quality stems from the usage of novel, comprehensive and custom-designed inputs that utilize information extracted from the sequence, evolutionary profiles, several sequence-predicted structural descriptors and sequence alignment. Analysis of the predictive model reveals several sequence-derived hallmarks of nucleotide-binding residues; they are usually conserved and flanked by less conserved residues, and they are associated with certain arrangements of secondary structures and amino acid pairs in the specific neighboring positions in the sequence. AVAILABILITY: http://biomine.ece.ualberta.ca/nSITEpred/ CONTACT: [email protected] SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: Molecular recognition features (MoRFs) are short binding regions located within longer intrinsically disordered regions that bind to protein partners via disorder-to-order transitions. MoRFs are implicated in important processes including signaling and regulation. However, only a limited number of experimentally validated MoRFs is known, which motivates development of computational methods that predict MoRFs from protein chains. RESULTS: We introduce a new MoRF predictor, MoRFpred, which identifies all MoRF types (α, β, coil and complex). We develop a comprehensive dataset of annotated MoRFs to build and empirically compare our method. MoRFpred utilizes a novel design in which annotations generated by sequence alignment are fused with predictions generated by a Support Vector Machine (SVM), which uses a custom designed set of sequence-derived features. The features provide information about evolutionary profiles, selected physiochemical properties of amino acids, and predicted disorder, solvent accessibility and B-factors. Empirical evaluation on several datasets shows that MoRFpred outperforms related methods: α-MoRF-Pred that predicts α-MoRFs and ANCHOR which finds disordered regions that become ordered when bound to a globular partner. We show that our predicted (new) MoRF regions have non-random sequence similarity with native MoRFs. We use this observation along with the fact that predictions with higher probability are more accurate to identify putative MoRF regions. We also identify a few sequence-derived hallmarks of MoRFs. They are characterized by dips in the disorder predictions and higher hydrophobicity and stability when compared to adjacent (in the chain) residues. AVAILABILITY: http://biomine.ece.ualberta.ca/MoRFpred/; http://biomine.ece.ualberta.ca/MoRFpred/Supplement.pdf.

Fuzzy cognitive maps (FCMs) are convenient and widely used architectures for modeling dynamic systems, which are characterized by a great deal of flexibility and adaptability. Several recent works in this area concern strategies for the development of FCMs. Although a few fully automated algorithms to learn these models from data have been introduced, the resulting FCMs are structurally considerably different than those developed by human experts. In particular, maps that were learned from data are much denser (with the density over 90% versus about 40% density of maps developed by humans). The sparseness of the maps is associated with their interpretability: the smaller the number of connections is, the higher is the transparency of the map. To this end, a novel learning approach, sparse real-coded genetic algorithms (SRCGAs), to learn FCMs is proposed. The method utilizes a density parameter to guide the learning toward a formation of maps of a certain predefined density. Comparative tests carried out for both synthetic and real-world data demonstrate that, given a suitable density estimate, the SRCGA method significantly outperforms other state-of-the-art learning methods. When the density estimate is unknown, the new method can be used in an automated fashion using a default value, and it is still able to produce models whose performance exceeds or is equal to the performance of the models generated by other methods.

Sequence-based prediction of protein secondary structure (SS) enjoys wide-spread and increasing use for the analysis and prediction of numerous structural and functional characteristics of proteins. The lack of a recent comprehensive and large-scale comparison of the numerous prediction methods results in an often arbitrary selection of a SS predictor. To address this void, we compare and analyze 12 popular, standalone and high-throughput predictors on a large set of 1975 proteins to provide in-depth, novel and practical insights. We show that there is no universally best predictor and thus detailed comparative studies are needed to support informed selection of SS predictors for a given application. Our study shows that the three-state accuracy (Q3) and segment overlap (SOV3) of the SS prediction currently reach 82% and 81%, respectively. We demonstrate that carefully designed consensus-based predictors improve the Q3 by additional 2% and that homology modeling-based methods are significantly better by 1.5% Q3 than ab initio approaches. Our empirical analysis reveals that solvent exposed and flexible coils are predicted with a higher quality than the buried and rigid coils, while inverse is true for the strands and helices. We also show that longer helices are easier to predict, which is in contrast to longer strands that are harder to find. The current methods confuse 1-6% of strand residues with helical residues and vice versa and they perform poorly for residues in the β- bridge and 3(10)-helix conformations. Finally, we compare predictions of the standalone implementations of four well-performing methods with their corresponding web servers.

MOTIVATION: X-ray crystallography-based protein structure determination, which accounts for majority of solved structures, is characterized by relatively low success rates. One solution is to build tools which support selection of targets that are more likely to crystallize. Several in silico methods that predict propensity of diffraction-quality crystallization from protein chains were developed. We show that the quality of their predictions drops when applied to more recent crystallization trails, which calls for new solutions. We propose a novel approach that alleviates drawbacks of the existing methods by using a recent dataset and improved protocol to annotate progress along the crystallization process, by predicting the success of the entire process and steps which result in the failed attempts, and by utilizing a compact and comprehensive set of sequence-derived inputs to generate accurate predictions. RESULTS: The proposed PPCpred (predictor of protein Production, Purification and Crystallization) predict propensity for production of diffraction-quality crystals, production of crystals, purification and production of the protein material. PPCpred utilizes comprehensive set of inputs based on energy and hydrophobicity indices, composition of certain amino acid types, predicted disorder, secondary structure and solvent accessibility, and content of certain buried and exposed residues. Our method significantly outperforms alignment-based predictions and several modern crystallization propensity predictors. Receiver operating characteristic (ROC) curves show that PPCpred is particularly useful for users who desire high true positive (TP) rates, i.e. low rate of mispredictions for solvable chains. Our model reveals several intuitive factors that influence the success of individual steps and the entire crystallization process, including the content of Cys, buried His and Ser, hydrophobic/hydrophilic segments and the number of predicted disordered segments. AVAILABILITY: http://biomine.ece.ualberta.ca/PPCpred/. CONTACT: [email protected].

BACKGROUND: Intrinsically disordered proteins play important roles in various cellular activities and their prevalence was implicated in a number of human diseases. The knowledge of the content of the intrinsic disorder in proteins is useful for a variety of studies including estimation of the abundance of disorder in protein families, classes, and complete proteomes, and for the analysis of disorder-related protein functions. The above investigations currently utilize the disorder content derived from the per-residue disorder predictions. We show that these predictions may over-or under-predict the overall amount of disorder, which motivates development of novel tools for direct and accurate sequence-based prediction of the disorder content. RESULTS: We hypothesize that sequence-level aggregation of input information may provide more accurate content prediction when compared with the content extracted from the local window-based residue-level disorder predictors. We propose a novel predictor, DisCon, that takes advantage of a small set of 29 custom-designed descriptors that aggregate and hybridize information concerning sequence, evolutionary profiles, and predicted secondary structure, solvent accessibility, flexibility, and annotation of globular domains. Using these descriptors and a ridge regression model, DisCon predicts the content with low, 0.05, mean squared error and high, 0.68, Pearson correlation. This is a statistically significant improvement over the content computed from outputs of ten modern disorder predictors on a test dataset with proteins that share low sequence identity with the training sequences. The proposed predictive model is analyzed to discuss factors related to the prediction of the disorder content. CONCLUSIONS: DisCon is a high-quality alternative for high-throughput annotation of the disorder content. We also empirically demonstrate that the DisCon's predictions can be used to improve binary annotations of the disordered residues from the real-value disorder propensities generated by current residue-level disorder predictors. The web server that implements the DisCon is available at http://biomine.ece.ualberta.ca/DisCon/.

ATP is a ubiquitous nucleotide that provides energy for cellular activities, catalyzes chemical reactions, and is involved in cellular signaling. The knowledge of the ATP-protein interactions helps with annotation of protein functions and finds applications in drug design. We propose a high-throughput machine learning-based predictor, ATPsite, which identifies ATP-binding residues from protein sequences. Statistical tests show that ATPsite significantly outperforms existing ATPint predictor and other solutions which utilize sequence alignment and residue conservation scoring. The improvements stem from the usage of novel custom-designed input features that are based on the sequence, evolutionary profiles, and the sequence-predicted structural descriptors including secondary structure, solvent accessibility, and dihedral angles. A simple consensus of the ATPsite with the sequence-alignment based predictor is shown to give further improvements.

Abstract Motivation: Intrinsically disordered proteins play a crucial role in numerous regulatory processes. Their abundance and ubiquity combined with a relatively low quantity of their annotations motivate research toward the development of computational models that predict disordered regions from protein sequences. Although the prediction quality of these methods continues to rise, novel and improved predictors are urgently needed. Results: We propose a novel method, named MFDp (Multilayered Fusion-based Disorder predictor), that aims to improve over the current disorder predictors. MFDp is as an ensemble of 3 Support Vector Machines specialized for the prediction of short, long and generic disordered regions. It combines three complementary disorder predictors, sequence, sequence profiles, predicted secondary structure, solvent accessibility, backbone dihedral torsion angles, residue flexibility and B-factors. Our method utilizes a custom-designed set of features that are based on raw predictions and aggregated raw values and recognizes various types of disorder. The MFDp is compared at the residue level on two datasets against eight recent disorder predictors and top-performing methods from the most recent CASP8 experiment. In spite of using training chains with ≤25% similarity to the test sequences, our method consistently and significantly outperforms the other methods based on the MCC index. The MFDp outperforms modern disorder predictors for the binary disorder assignment and provides competitive real-valued predictions. The MFDp's outputs are also shown to outperform the other methods in the identification of proteins with long disordered regions. Availability: http://biomine.ece.ualberta.ca/MFDp.html Supplementary information: Supplementary data are available at Bioinformatics online. Contact: [email protected]

In this paper, we introduce a novel system for recognition of partially occluded and rotated images. The system is based on a hierarchical network of integrate-and-fire spiking neurons with random synaptic connections and a novel organization process. The network generates integrated output sequences that are used for image classification. The proposed network is shown to provide satisfactory predictive performance given that the number of the recognition neurons and synaptic connections are adjusted to the size of the input image. Comparison of synaptic plasticity activity rule (SAPR) and spike timing dependant plasticity rules, which are used to learn connections between the spiking neurons, indicates that the former gives better results and thus the SAPR rule is used. Test results show that the proposed network performs better than a recognition system based on support vector machines.

BACKGROUND: Knowledge of structural class is used by numerous methods for identification of structural/functional characteristics of proteins and could be used for the detection of remote homologues, particularly for chains that share twilight-zone similarity. In contrast to existing sequence-based structural class predictors, which target four major classes and which are designed for high identity sequences, we predict seven classes from sequences that share twilight-zone identity with the training sequences. RESULTS: The proposed MODular Approach to Structural class prediction (MODAS) method is unique as it allows for selection of any subset of the classes. MODAS is also the first to utilize a novel, custom-built feature-based sequence representation that combines evolutionary profiles and predicted secondary structure. The features quantify information relevant to the definition of the classes including conservation of residues and arrangement and number of helix/strand segments. Our comprehensive design considers 8 feature selection methods and 4 classifiers to develop Support Vector Machine-based classifiers that are tailored for each of the seven classes. Tests on 5 twilight-zone and 1 high-similarity benchmark datasets and comparison with over two dozens of modern competing predictors show that MODAS provides the best overall accuracy that ranges between 80% and 96.7% (83.5% for the twilight-zone datasets), depending on the dataset. This translates into 19% and 8% error rate reduction when compared against the best performing competing method on two largest datasets. The proposed predictor provides accurate predictions at 58% accuracy for membrane proteins class, which is not considered by majority of existing methods, in spite that this class accounts for only 2% of the data. Our predictive model is analyzed to demonstrate how and why the input features are associated with the corresponding classes. CONCLUSIONS: The improved predictions stem from the novel features that express collocation of the secondary structure segments in the protein sequence and that combine evolutionary and secondary structure information. Our work demonstrates that conservation and arrangement of the secondary structure segments predicted along the protein chain can successfully predict structural classes which are defined based on the spatial arrangement of the secondary structures. A web server is available at http://biomine.ece.ualberta.ca/MODAS/.

Fuzzy cognitive maps (FCMs) are a convenient tool for modeling of dynamic systems by means of concepts connected by cause-effect relationships. The FCM models can be developed either manually (by the experts) or using an automated learning method (from data). Some of the methods from the latter group, including recently proposed Nonlinear Hebbian Learning (NHL) algorithm, use Hebbian law and a set of conditions imposed on output concepts. In this paper, we propose a novel approach named data-driven NHL (DD-NHL) that extends NHL method by using historical data of the input concepts to provide improved quality of the learned FCMs. DD-NHL is tested on both synthetic and real-life data, and the experiments show that if historical data are available, then the proposed method produces better FCM models when compared with those formed by the generic NHL method.

While data could be discrete and continuous (defined as ordinal numerical features), some classifiers, like Naive Bayes (NB), work only with or may perform better with the discrete data. We focus on NB due to its popularity and linear training time. We investigate the impact of eight discretization algorithms (Equal Width, Equal Frequency, Maximum Entropy, IEM, CADD, CAIM, MODL, and CACC) on the classification with NB and two modern semi-NB classifiers, LBR and AODE.Our comprehensive empirical study indicates that unsupervised discretization algorithms are the fastest while among the supervised algorithms the fastest is maximum entropy, followed by CAIM and IEM. The CAIM and MODL discretizers generate the lowest and the highest number of discrete values, respectively.We compare the time to build the classification model and classification accuracy when using raw and discretized data. We show that discretization helps to improve the classification with the NB when compared with flexible NB which models continuous features using Gaussian kernels. The AODE classifier obtains on average the best accuracy, while the best performing setup includes discretization with IEM and classification with AODE. The runner-up setups include CAIM and CACC coupled with AODE and CAIM and IEM coupled with LBR. IEM and CAIM are shown to provide statistically significant improvements across all considered datasets for LBR and AODE classifiers when compared with using NB on the continuous data. We also show that the improved accuracy comes at the trade-off of substantially increased runtime.

MOTIVATION: Prediction of catalytic residues provides useful information for the research on function of enzymes. Most of the existing prediction methods are based on structural information, which limits their use. We propose a sequence-based catalytic residue predictor that provides predictions with quality comparable to modern structure-based methods and that exceeds quality of state-of-the-art sequence-based methods. RESULTS: Our method (CRpred) uses sequence-based features and the sequence-derived PSI-BLAST profile. We used feature selection to reduce the dimensionality of the input (and explain the input) to support vector machine (SVM) classifier that provides predictions. Tests on eight datasets and side-by-side comparison with six modern structure- and sequence-based predictors show that CRpred provides predictions with quality comparable to current structure-based methods and better than sequence-based methods. The proposed method obtains 15-19% precision and 48-58% TP (true positive) rate, depending on the dataset used. CRpred also provides confidence values that allow selecting a subset of predictions with higher precision. The improved quality is due to newly designed features and careful parameterization of the SVM. The features incorporate amino acids characterized by the highest and the lowest propensities to constitute catalytic residues, Gly that provides flexibility for catalytic sites and sequence motifs characteristic to certain catalytic reactions. Our features indicate that catalytic residues are on average more conserved when compared with the general population of residues and that highly conserved amino acids characterized by high catalytic propensity are likely to form catalytic sites. We also show that local (with respect to the sequence) hydrophobicity contributes towards the prediction.

BACKGROUND: Accurate peptide identification is important to high-throughput proteomics analyses that use mass spectrometry. Search programs compare fragmentation spectra (MS/MS) of peptides from complex digests with theoretically derived spectra from a database of protein sequences. Improved discrimination is achieved with theoretical spectra that are based on simulating gas phase chemistry of the peptides, but the limited understanding of those processes affects the accuracy of predictions from theoretical spectra. RESULTS: We employed a robust data mining strategy using new feature annotation functions of MAE software, which revealed under-prediction of the frequency of occurrence in fragmentation of the second peptide bond. We applied methods of exploratory data analysis to pre-process the information in the MS/MS spectra, including data normalization and attribute selection, to reduce the attributes to a smaller, less correlated set for machine learning studies. We then compared our rule building machine learning program, DataSqueezer, with commonly used association rules and decision tree algorithms. All used machine learning algorithms produced similar results that were consistent with expected properties for a second gas phase mechanism at the second peptide bond. CONCLUSION: The results provide compelling evidence that we have identified underlying chemical properties in the data that suggest the existence of an additional gas phase mechanism for the second peptide bond. Thus, the methods described in this study provide a valuable approach for analyses of this kind in the future.

BACKGROUND: Protein structure prediction methods provide accurate results when a homologous protein is predicted, while poorer predictions are obtained in the absence of homologous templates. However, some protein chains that share twilight-zone pairwise identity can form similar folds and thus determining structural similarity without the sequence similarity would be desirable for the structure prediction. The folding type of a protein or its domain is defined as the structural class. Current structural class prediction methods that predict the four structural classes defined in SCOP provide up to 63% accuracy for the datasets in which sequence identity of any pair of sequences belongs to the twilight-zone. We propose SCPRED method that improves prediction accuracy for sequences that share twilight-zone pairwise similarity with sequences used for the prediction. RESULTS: SCPRED uses a support vector machine classifier that takes several custom-designed features as its input to predict the structural classes. Based on extensive design that considers over 2300 index-, composition- and physicochemical properties-based features along with features based on the predicted secondary structure and content, the classifier's input includes 8 features based on information extracted from the secondary structure predicted with PSI-PRED and one feature computed from the sequence. Tests performed with datasets of 1673 protein chains, in which any pair of sequences shares twilight-zone similarity, show that SCPRED obtains 80.3% accuracy when predicting the four SCOP-defined structural classes, which is superior when compared with over a dozen recent competing methods that are based on support vector machine, logistic regression, and ensemble of classifiers predictors. CONCLUSION: The SCPRED can accurately find similar structures for sequences that share low identity with sequence used for the prediction. The high predictive accuracy achieved by SCPRED is attributed to the design of the features, which are capable of separating the structural classes in spite of their low dimensionality. We also demonstrate that the SCPRED's predictions can be successfully used as a post-processing filter to improve performance of modern fold classification methods.

BACKGROUND: Residue depth allows determining how deeply a given residue is buried, in contrast to the solvent accessibility that differentiates between buried and solvent-exposed residues. When compared with the solvent accessibility, the depth allows studying deep-level structures and functional sites, and formation of the protein folding nucleus. Accurate prediction of residue depth would provide valuable information for fold recognition, prediction of functional sites, and protein design. RESULTS: A new method, RDPred, for the real-value depth prediction from protein sequence is proposed. RDPred combines information extracted from the sequence, PSI-BLAST scoring matrices, and secondary structure predicted with PSIPRED. Three-fold/ten-fold cross validation based tests performed on three independent, low-identity datasets show that the distance based depth (computed using MSMS) predicted by RDPred is characterized by 0.67/0.67, 0.66/0.67, and 0.64/0.65 correlation with the actual depth, by the mean absolute errors equal 0.56/0.56, 0.61/0.60, and 0.58/0.57, and by the mean relative errors equal 17.0%/16.9%, 18.2%/18.1%, and 17.7%/17.6%, respectively. The mean absolute and the mean relative errors are shown to be statistically significantly better when compared with a method recently proposed by Yuan and Wang [Proteins 2008; 70:509-516]. The results show that three-fold cross validation underestimates the variability of the prediction quality when compared with the results based on the ten-fold cross validation. We also show that the hydrophilic and flexible residues are predicted more accurately than hydrophobic and rigid residues. Similarly, the charged residues that include Lys, Glu, Asp, and Arg are the most accurately predicted. Our analysis reveals that evolutionary information encoded using PSSM is characterized by stronger correlation with the depth for hydrophilic amino acids (AAs) and aliphatic AAs when compared with hydrophobic AAs and aromatic AAs. Finally, we show that the secondary structure of coils and strands is useful in depth prediction, in contrast to helices that have relatively uniform distribution over the protein depth. Application of the predicted residue depth to prediction of buried/exposed residues shows consistent improvements in detection rates of both buried and exposed residues when compared with the competing method. Finally, we contrasted the prediction performance among distance based (MSMS and DPX) and volume based (SADIC) depth definitions. We found that the distance based indices are harder to predict due to the more complex nature of the corresponding depth profiles. CONCLUSION: The proposed method, RDPred, provides statistically significantly better predictions of residue depth when compared with the competing method. The predicted depth can be used to provide improved prediction of both buried and exposed residues. The prediction of exposed residues has implications in characterization/prediction of interactions with ligands and other proteins, while the prediction of buried residues could be used in the context of folding predictions and simulations.

BACKGROUND: beta-turn is a secondary protein structure type that plays significant role in protein folding, stability, and molecular recognition. To date, several methods for prediction of beta-turns from protein sequences were developed, but they are characterized by relatively poor prediction quality. The novelty of the proposed sequence-based beta-turn predictor stems from the usage of a window based information extracted from four predicted three-state secondary structures, which together with a selected set of position specific scoring matrix (PSSM) values serve as an input to the support vector machine (SVM) predictor. RESULTS: We show that (1) all four predicted secondary structures are useful; (2) the most useful information extracted from the predicted secondary structure includes the structure of the predicted residue, secondary structure content in a window around the predicted residue, and features that indicate whether the predicted residue is inside a secondary structure segment; (3) the PSSM values of Asn, Asp, Gly, Ile, Leu, Met, Pro, and Val were among the top ranked features, which corroborates with recent studies. The Asn, Asp, Gly, and Pro indicate potential beta-turns, while the remaining four amino acids are useful to predict non-beta-turns. Empirical evaluation using three nonredundant datasets shows favorable Q total, Q predicted and MCC values when compared with over a dozen of modern competing methods. Our method is the first to break the 80% Q total barrier and achieves Q total = 80.9%, MCC = 0.47, and Q predicted higher by over 6% when compared with the second best method. We use feature selection to reduce the dimensionality of the feature vector used as the input for the proposed prediction method. The applied feature set is smaller by 86, 62 and 37% when compared with the second and two third-best (with respect to MCC) competing methods, respectively. CONCLUSION: Experiments show that the proposed method constitutes an improvement over the competing prediction methods. The proposed prediction model can better discriminate between beta-turns and non-beta-turns due to obtaining lower numbers of false positive predictions. The prediction model and datasets are freely available at http://biomine.ece.ualberta.ca/BTNpred/BTNpred.html.

In this paper, we introduce a novel approach to time-series prediction realized both at the linguistic and numerical level. It exploits fuzzy cognitive maps (FCMs) along with a recently proposed learning method that takes advantage of real-coded genetic algorithms. FCMs are used for modeling and qualitative analysis of dynamic systems. Within the framework of FCMs, the systems are described by means of concepts and their mutual relationships. The proposed prediction method combines FCMs with granular, fuzzy-set-based model of inputs. One of their main advantages is an ability to carry out modeling and prediction at both numerical and linguistic levels. A comprehensive set of experiments has been carried out with two major goals in mind. One is to assess quality of the proposed architecture, the other to examine the influence of its parameters of the prediction technique on the quality of prediction. The obtained results, which are compared with other prediction techniques using fuzzy sets, demonstrate that the proposed architecture offers substantial accuracy expressed at both linguistic and numerical levels.

Fuzzy cognitive maps (FCMs) are a class of discrete-time artificial neural networks that are used to model dynamic systems. A recently introduced supervised learning method, which is based on real-coded genetic algorithm (RCGA), allows learning high-quality FCMs from historical data. The current bottleneck of this learning method is its scalability, which originates from large continuous search space (of quadratic size with respect to the size of the FCM) and computational complexity of genetic optimization. To this end, the goal of this paper is to explore parallel nature of genetic algorithms to alleviate the scalability problem. We use the global single-population master-slave parallelization method to speed up the FCMs learning method. We investigate the influence of different hardware architectures on the computational time of the learning method by executing a wide range of synthetic and real-life benchmarking tests. We analyze the quality of the proposed parallel learning method in application to both dense and sparse large FCMs, i.e. maps that consist of several dozens of concepts. The parallelization is shown to provide substantial speed-ups, allowing doubling the size of the FCM that can be learned by parallelization with 8 processors.

MOTIVATION: The number of protein families has been estimated to be as small as 1000. Recent study shows that the growth in discovery of novel structures that are deposited into PDB and the related rate of increase of SCOP categories are slowing down. This indicates that the protein structure space will be soon covered and thus we may be able to derive most of remaining structures by using the known folding patterns. Present tertiary structure prediction methods behave well when a homologous structure is predicted, but give poorer results when no homologous templates are available. At the same time, some proteins that share twilight-zone sequence identity can form similar folds. Therefore, determination of structural similarity without sequence similarity would be beneficial for prediction of tertiary structures. RESULTS: The proposed PFRES method for automated protein fold classification from low identity (<35%) sequences obtains 66.4% and 68.4% accuracy for two test sets, respectively. PFRES obtains 6.3-12.4% higher accuracy than the existing methods. The prediction accuracy of PFRES is shown to be statistically significantly better than the accuracy of competing methods. Our method adopts a carefully designed, ensemble-based classifier, and a novel, compact and custom-designed feature representation that includes nearly 90% less features than the representation of the most accurate competing method (36 versus 283). The proposed representation combines evolutionary information by using the PSI-BLAST profile-based composition vector and information extracted from the secondary structure predicted with PSI-PRED. AVAILABILITY: The method is freely available from the authors upon request.

Many of the industrial and research databases are plagued by the problem of missing values. Some evident examples include databases associated with instrument maintenance, medical applications, and surveys. One of the common ways to cope with missing values is to complete their imputation (filling in). Given the rapid growth of sizes of databases, it becomes imperative to come up with a new imputation methodology along with efficient algorithms. The main objective of this paper is to develop a unified framework supporting a host of imputation methods. In the development of this framework, we require that its usage should (on average) lead to the significant improvement of accuracy of imputation while maintaining the same asymptotic computational complexity of the individual methods. Our intent is to provide a comprehensive review of the representative imputation techniques. It is noticeable that the use of the framework in the case of a low-quality single-imputation method has resulted in the imputation accuracy that is comparable to the one achieved when dealing with some other advanced imputation techniques. We also demonstrate, both theoretically and experimentally, that the application of the proposed framework leads to a linear computational complexity and, therefore, does not affect the asymptotic complexity of the associated imputation method.

Sliding window based methods are relatively often applied in prediction of various aspects related to protein structure. Despite their wide spread use, researchers did not establish a standard related to the size of the window, i.e., window sizes ranging between 7 and 17 residues were used in the past. To this end, this paper performs a computational study based on a probabilistic approach that aims at finding an optimal sliding window size. The results shows that formation of helical structure can be affected by amino acids (AAs) that are up to 9 positions away in the sequence, while the formation of coils and strands can be affected by AAs that are up to 3 and 6 positions away, respectively. Overall, our results suggest that a sliding window with 19 residues is optimal for secondary structure prediction, while for a specific prediction tasks, such as prediction of p-strands, a smaller window size is sufficient. Finally, the 20 AAs are categorized into five groups based on their influence of formation of the secondary structure. The finding related to the optimal window size was confirmed based on an independent experimental study related to the prediction of secondary protein structure

Protein structure prediction is one of the core research areas in bioinformatics. This paper addresses the protein secondary structure prediction problem for the twilight zone proteins, which are characterized by low, about 25% homology to the sets of known sequences. The commonly used sequence alignment based algorithms fail to provide accurate prediction for sequences of such low homology, and thus alternative solutions should be sought. We propose a novel method that aims at the prediction of the number of helical structures based on the twilight zone protein sequences. The method is based on a custom designed and compact feature based sequences representation and applies a decision tree prediction algorithm. The performed experimental study shows superiority of the proposed method over three other prediction algorithms and the results provided by YASPIN algorithm, which is a state-of-the-art alignment based secondary structure prediction method designed using low homology sequences

This paper addresses in silico prediction of protein structural classes as defined in the SCOP database. The SCOP defines total of 11 classes, while majority of proteins are classified to the 4 classes: all-alpha all-beta alpha/beta, and alpha+beta. The main goals of this paper are to experimentally evaluate the impact of predicted protein secondary structure content on the structural class prediction and to develop a novel protein sequence representation. The experiments include application of three protein sequence representations and four classifiers to prediction of both 4 and 11 structural classes. The predictions are performed using a large dataset of low homology (twilight zone) sequences. The proposed sequence representation includes the predicted structural content, which provides the strongest contribution towards classification, composition and composition moment vectors, hydrophobic autocorrelations, chemical group composition and molecular weight of the protein. The predicted content values are shown on average to improve the prediction accuracy by 3.3% and 4.2% for the 4 and 11 classes, respectively, when compared to sequence representation that does not utilize this information. Finally, we propose a very compact, 20 dimensional sequence representation that is shown to improve the prediction accuracy by 5.1-8.5% when compared with recently published results

In this paper, we describe a new Synaptic Plasticity Activity Rule (SAPR) developed for use in networks of spiking neurons. Such networks can be used for simulations of physiological experiments as well as for other computations like image analysis. Most synaptic plasticity rules use artificially defined functions to modify synaptic connection strengths. In contrast, our rule makes use of the existing postsynaptic potential values to compute the value of adjustment. The network of spiking neurons we consider consists of excitatory and inhibitory neurons. Each neuron is implemented as an integrate-and-fire model that accurately mimics the behavior of biological neurons. To test performance of our new plasticity rule we designed a model of a biologically-inspired signal processing system, and used it for object detection in eye images of diabetic retinopathy patients, and lung images of cystic fibrosis patients. The results show that the network detects the edges of objects within an image, essentially segmenting it. Our ultimate goal, however, is not the development of an image segmentation tool that would be more efficient than nonbiological algorithms, but developing a physiologically correct neural network model that could be applied to a wide range of neurological experiments. We decided to validate the SAPR by using it in a network of spiking neurons for image segmentation because it is easy to visually assess the results. An important thing is that image segmentation is done in an entirely unsupervised way.

Business intelligence and bioinformatics applications increasingly require the mining of datasets consisting of millions of data points, or crafting real-time enterprise-level decision support systems for large corporations and drug companies. In all cases, there needs to be an underlying data mining system, and this mining system must be highly scalable. To this end, we describe a new rule learner called DataSqueezer. The learner belongs to the family of inductive supervised rule extraction algorithms. DataSqueezer is a simple, greedy, rule builder that generates a set of production rules from labeled input data. In spite of its relative simplicity, DataSqueezer is a very effective learner. The rules generated by the algorithm are compact, comprehensible, and have accuracy comparable to rules generated by other state-of-the-art rule extraction algorithms. The main advantages of DataSqueezer are very high efficiency, and missing data resistance. DataSqueezer exhibits log-linear asymptotic complexity with the number of training examples, and it is faster than other state-of-the-art rule learners. The learner is also robust to large quantities of missing data, as verified by extensive experimental comparison with the other learners. DataSqueezer is thus well suited to modern data mining and business intelligence tasks, which commonly involve huge datasets with a large fraction of missing data.

Fuzzy cognitive maps (FCMs) form a convenient, simple, and powerful tool for simulation and analysis of dynamic systems. The popularity of FCMs stems from their simplicity and transparency. While being successful in a variety of application domains, FCMs are hindered by necessity of involving domain experts to develop the model. Since human experts are subjective and can handle only relatively simple networks (maps), there is an urgent need to develop methods for automated generation of FCM models. This study proposes a novel evolutionary learning that is able to generate FCM models from input historical data, and without any human intervention. The proposed method is based on genetic algorithms, and is carried out through supervised learning. The paper tests the method through a series of carefully selected experimental studies

Ability to provide convenient access to scientific documents becomes a difficult problem due to large and constantly increasing number of incoming documents and extensive manual work associated with their storage, description and classification. This requires intelligent search and classification capabilities for users to find required information. It is especially true for repositories of scientific medical articles due to their extensive use, large size and number of new documents, and well maintained structure. This research aims to provide an automated method for classification of articles into the structure of medical document repositories, which would support currently performed extensive manual work. The proposed method classifies articles from the largest medical repository, MEDLINE, using state of the art data mining technology. The method is based on a novel associative classification technique which considers recurrent items and most importantly multi-label characteristic of the MEDLINE data. Based on large scale experiments that utilize 350,000 documents several different classification algorithms have been compared including both recurrent and non-recurrent associative classification. The algorithms are capable of assigning each medical document to several classes (multi-label classification) and are characterized by relatively high accuracy. We also investigate different measures of classification quality and point out pros and cons of each. Based on experimental result we show that recurrent item based associative classification demonstrates superior performance and propose three alternative setups that allow the user to obtain different desired classification qualities.

Complexity, or in other words compactness, of models generated by rule learners is one of often neglected issues, although it has a profound effect on the success of any project that utilizes the rules. Researchers strive to propose learners that are characterized by excellent accuracy, and sometimes also low computational complexity, but the size of the data model generated by the learners is often not even reported. While the model size can be disregarded from the research point of view, it is very important from the end user's perspective. Quite often the generated model is too complex to be manually analyzed or inspected, which prohibits from using it in a real-world setting. To jill this gap, the paper proposes a novel framework, which is designed to address problem of complexity reduction of rule based models. The framework is based on a Meta Mining concept, and can be applied to enhance several of existing rule learners. Its main goal is to reduce complexity, in terms of reducing size and number of generated rules, without sacrificing accuracy of the rules. The paper proposes the framework, and tests it on a set of benchmark datasets using two well known rule learners: C5.0 and DataSqueezer. The results are encouraging, and show that 50% complexity reduction can be achieved virtually without any loss of accuracy.

The task of extracting knowledge from databases is quite often performed by machine learning algorithms. The majority of these algorithms can be applied only to data described by discrete numerical or nominal attributes (features). In the case of continuous attributes, there is a need for a discretization algorithm that transforms continuous attributes into discrete ones. We describe such an algorithm, called CAIM (class-attribute interdependence maximization), which is designed to work with supervised data. The goal of the CAIM algorithm is to maximize the class-attribute interdependence and to generate a (possibly) minimal number of discrete intervals. The algorithm does not require the user to predefine the number of intervals, as opposed to some other discretization algorithms. The tests performed using CAIM and six other state-of-the-art discretization algorithms show that discrete attributes generated by the CAIM algorithm almost always have the lowest number of intervals and the highest class-attribute interdependency. Two machine learning algorithms, the CLIP4 rule algorithm and the decision tree algorithm, are used to generate classification rules from data discretized by CAIM. For both the CLIP4 and decision tree algorithms, the accuracy of the generated rules is higher and the number of the rules is lower for data discretized using the CAIM algorithm when compared to data discretized using six other discretization algorithms. The highest classification accuracy was achieved for data sets discretized with the CAIM algorithm, as compared with the other six algorithms.