BACKGROUND: Regulation of gene expression, protein synthesis, replication and assembly of many viruses involve RNA-protein interactions. Although some successful computational tools have been reported to recognize RNA binding sites in proteins, the problem of specificity remains poorly investigated. After the nucleotide base composition, the dinucleotide is the smallest unit of RNA sequence information and many RNA-binding proteins simply bind to regions enriched in one dinucleotide. Interaction preferences of protein subsequences and dinucleotides can be inferred from protein-RNA complex structures, enabling a training-based prediction approach. RESULTS: We analyzed basic statistics of amino acid-dinucleotide contacts in protein-RNA complexes and found their pairing preferences could be identified. Using a standard approach to represent protein subsequences by their evolutionary profile, we trained neural networks to predict multiclass target vectors corresponding to 16 possible contacting dinucleotide subsequences. In the cross-validation experiments, the accuracies of the optimum network, measured as areas under the curve (AUC) of the receiver operating characteristic (ROC) graphs, were in the range of 65-80%. CONCLUSIONS: Dinucleotide-specific contact predictions have also been extended to the prediction of interacting protein and RNA fragment pairs, which shows the applicability of this method to predict targets of RNA-binding proteins. A web server predicting the 16-dimensional contact probability matrix directly from a user-defined protein sequence was implemented and made available at: http://tardis.nibio.go.jp/netasa/srcpred.

BACKGROUND: Many structural properties such as solvent accessibility, dihedral angles and helix-helix contacts can be assigned to each residue in a membrane protein. Independent studies exist on the analysis and sequence-based prediction of some of these so-called one-dimensional features. However, there is little explanation of why certain residues are predicted in a wrong structural class or with large errors in the absolute values of these features. On the other hand, membrane proteins undergo conformational changes to allow transport as well as ligand binding. These conformational changes often occur via residues that are inherently flexible and hence, predicting fluctuations in residue positions is of great significance. RESULTS: We performed a statistical analysis of common patterns among selected one-dimensional equilibrium structural features (ESFs) and developed a method for simultaneously predicting all of these features using an integrated system. Our results show that the prediction performance can be improved if multiple structural features are trained in an integrated model, compared to the current practice of developing individual models. In particular, the performance of the solvent accessibility and bend-angle prediction improved in this way. The well-performing bend-angle prediction can be used to predict helical positions with severe kinks at a modest success rate. Further, we showed that single-chain conformational dynamics, measured by B-factors derived from normal mode analysis, could be predicted from observed and predicted ESFs with good accuracy. A web server was developed (http://tardis.nibio.go.jp/netasa/htmone/) for predicting the one-dimensional ESFs from sequence information and analyzing the differences between the predicted and observed values of the ESFs. CONCLUSIONS: The prediction performance of the integrated model is significantly better than that of the models performing the task separately for each feature for the solvent accessibility and bend-angle predictions. The predictability of the features also plays a role in determining flexible positions. Although the dynamics studied here concerns local atomic fluctuations, a similar analysis in terms of global structural features will be helpful in predicting large-scale conformational changes, for which work is in progress.

Genetic algorithm (GA) applied to feature selection and model optimization improved the performance of robust mathematical models such as Bayesian-regularized neural networks (BRANNs) and support vector machines (SVMs) on different drug design datasets. The selection of optimum input variables and the adjustment of network weights and biases to optimum values to yield generalizable predictors were optimized by combining Bayesian training and GA based-variable selection. Similarly, kernel and regularization parameters of SVMs were properly set by GA optimization. The predictors were more accurate and robust than previous published models on the same datasets. In addition, feature selection over large pools of molecular descriptors allowed determining the structural and atomic properties of the ligands that are ruling the biological interactions with the target.

Similar neural network models based on single sequence and evolutionary profiles of residues have been successfully used in the past for predicting secondary structure, solvent accessibility, protein-, DNA- and carbohydrate- binding sites. ATP is a ubiquitous ligand in all living-systems, involved in most biological functions requiring energy and charge transfer. Prediction of ATP-binding site from single sequences and their evolutionary profiles at a high throughput rate can be used at genomic level as well as quick clues for site-directed mutagenesis experiments. We have developed a method for such predictions to demonstrate yet another application of sequence-base prediction algorithms using neural networks. This method can achieve 81% sensitivity and 69% specificity which are mutually adjustable in a wide range on a three-fold cross-validation data set.

We have utilized neural networks in different applications of bioinformatics such as discrimination of beta-barrel membrane proteins, mesophilic and thermophilic proteins, different folding types of globular proteins, different classes of transporter proteins and predicting the secondary structures of beta-barrel membrane proteins. In these methods, we have used the information about amino acid composition, neighboring residue information, inter-residue contacts and amino acid properties as features. We observed that the performance with neural networks is comparable to or better than other widely used machine learning techniques.

BACKGROUND: Detection of DNA-binding sites in proteins is of enormous interest for technologies targeting gene regulation and manipulation. We have previously shown that a residue and its sequence neighbor information can be used to predict DNA-binding candidates in a protein sequence. This sequence-based prediction method is applicable even if no sequence homology with a previously known DNA-binding protein is observed. Here we implement a neural network based algorithm to utilize evolutionary information of amino acid sequences in terms of their position specific scoring matrices (PSSMs) for a better prediction of DNA-binding sites. RESULTS: An average of sensitivity and specificity using PSSMs is up to 8.7% better than the prediction with sequence information only. Much smaller data sets could be used to generate PSSM with minimal loss of prediction accuracy. CONCLUSION: One problem in using PSSM-derived prediction is obtaining lengthy and time-consuming alignments against large sequence databases. In order to speed up the process of generating PSSMs, we tried to use different reference data sets (sequence space) against which a target protein is scanned for PSI-BLAST iterations. We find that a very small set of proteins can actually be used as such a reference data without losing much of the prediction value. This makes the process of generating PSSMs very rapid and even amenable to be used at a genome level. A web server has been developed to provide these predictions of DNA-binding sites for any new protein from its amino acid sequence. AVAILABILITY: Online predictions based on this method are available at http://www.netasa.org/dbs-pssm/

MOTIVATION: Though vitally important to cell function, the mechanism of protein-DNA binding has not yet been completely understood. We therefore analysed the relationship between DNA binding and protein sequence composition, solvent accessibility and secondary structure. Using non-redundant databases of transcription factors and protein-DNA complexes, neural network models were developed to utilize the information present in this relationship to predict DNA-binding proteins and their binding residues. RESULTS: Sequence composition was found to provide sufficient information to predict the probability of its binding to DNA with nearly 69% sensitivity at 64% accuracy for the considered proteins; sequence neighbourhood and solvent accessibility information were sufficient to make binding site predictions with 40% sensitivity at 79% accuracy. Detailed analysis of binding residues shows that some three- and five-residue segments frequently bind to DNA and that solvent accessibility plays a major role in binding. Although, binding behaviour was not associated with any particular secondary structure, there were interesting exceptions at the residue level. Over-representation of some residues in the binding sites was largely lost at the total sequence level, but a different kind of compositional preference was observed in DNA-binding proteins.

BACKGROUND: Accessible surface area (ASA) or solvent accessibility of amino acids in a protein has important implications. Knowledge of surface residues helps in locating potential candidates of active sites. Therefore, a method to quickly see the surface residues in a two dimensional model would help to immediately understand the population of amino acid residues on the surface and in the inner core of the proteins. RESULTS: ASAView is an algorithm, an application and a database of schematic representations of solvent accessibility of amino acid residues within proteins. A characteristic two-dimensional spiral plot of solvent accessibility provides a convenient graphical view of residues in terms of their exposed surface areas. In addition, sequential plots in the form of bar charts are also provided. Online plots of the proteins included in the entire Protein Data Bank (PDB), are provided for the entire protein as well as their chains separately. CONCLUSIONS: These graphical plots of solvent accessibility are likely to provide a quick view of the overall topological distribution of residues in proteins. Chain-wise computation of solvent accessibility is also provided.

SUMMARY: RVP-net is an online program for the prediction of real valued solvent accessibility. All previous methods of accessible surface area (ASA) predictions classify amino acid residues into exposure states and named them buried or exposed based on different thresholds. Real values in some cases were generated by taking the mid points of these state thresholds. This is the first method, which provides a direct prediction of ASA without making exposure categories and achieves results better than 19% mean absolute error. To facilitate batch processing of several sequences, a standalone version of this tool is also provided. AVAILABILITY: Online predictions are available at http://www.netasa.org/rvp-net/. Standalone version of the program can be obtained from the corresponding author by E-mail request.

Abstract Motivation: Prediction of the tertiary structure of a protein from its amino acid sequence is one of the most important problems in molecular biology. The successful prediction of solvent accessibility will be very helpful to achieve this goal. In the present work, we have implemented a server, NETASA for predicting solvent accessibility of amino acids using our newly optimized neural network algorithm. Several new features in the neural network architecture and training method have been introduced, and the network learns faster to provide accuracy values, which are comparable or better than other methods of ASA prediction. Results: Prediction in two and three state classification systems with several thresholds are provided. Our prediction method achieved the accuracy level upto 90% for training and 88% for test data sets. Three state prediction results provide a maximum 65% accuracy for training and 63% for the test data. Applicability of neural networks for ASA prediction has been confirmed with a larger data set and wider range of state thresholds. Salient differences between a linear and exponential network for ASA prediction have been analysed. Availability: Online predictions are freely available at: http://www.netasa.org. Linux ix86 binaries of the program written for this work may be obtained by email from the corresponding author. Contact: [email protected] * To whom correspondence should be addressed at: RIKEN Tsukuba Institute, 3-1-1, Koyadai, Tsukuba 305 0074, Ibaraki, Japan. † Present address: Computational Biology Research Center (CBRC), AIST 2-41-6 Aomi, Koto-ku, Tokyo 135-0064, Japan.