Cross-validation of protein structural class prediction using statistical clustering and neural networks.

We present an approach to predicting protein structural class that uses amino acid composition and hydrophobic pattern frequency information as input to two types of neural networks: (1) a three-layer back-propagation network and (2) a learning vector quantization network. The results of these methods are compared to those obtained from a modified Euclidean statistical clustering algorithm. The protein sequence data used to drive these algorithms consist of the normalized frequency of up to 20 amino acid types and six hydrophobic amino acid patterns. From these frequency values the structural class predictions for each protein (all-alpha, all-beta, or alpha-beta classes) are derived. Examples consisting of 64 previously classified proteins were randomly divided into multiple training (56 proteins) and test (8 proteins) sets. The best performing algorithm on the test sets was the learning vector quantization network using 17 inputs, obtaining a prediction accuracy of 80.2%. The Matthews correlation coefficients are statistically significant for all algorithms and all structural classes. The differences between algorithms are in general not statistically significant. These results show that information exists in protein primary sequences that is easily obtainable and useful for the prediction of protein structural class by neural networks as well as by standard statistical clustering algorithms.     

A hybrid genetic-neural model for predicting protein structural classes

A genetic algorithm (GA) for feature selection in conjunction with neural network was applied to predict protein structural classes based on single amino acid and all dipeptide composition frequencies. These sequence parameters were encoded as input features for a GA in feature selection procedure and classified with a three-layered neural network to predict protein structural classes. The system was established through optimization of the classification performance of neural network which was used as evaluation function. In this study, self-consistency and jackknife tests on a database containing 498 proteins were used to verify the performance of this hybrid method, and were compared with some of prior works. The adoption of a hybrid model, which encompasses genetic and neural technologies, demonstrated to be a promising approach in the task of protein structural class prediction.     

Predicting membrane protein types by fusing composite protein sequence features into pseudo amino acid composition

Membrane proteins are vital type of proteins that serve as channels, receptors, and energy transducers in a cell. Prediction of membrane protein types is an important research area in bioinformatics. Knowledge of membrane protein types provides some valuable information for predicting novel example of the membrane protein types. However, classification of membrane protein types can be both time consuming and susceptible to errors due to the inherent similarity of membrane protein types. In this paper, neural networks based membrane protein type prediction system is proposed. Composite protein sequence representation (CPSR) is used to extract the features of a protein sequence, which includes seven feature sets; amino acid composition, sequence length, 2 gram exchange group frequency, hydrophobic group, electronic group, sum of hydrophobicity, and R-group. Principal component analysis is then employed to reduce the dimensionality of the feature vector. The probabilistic neural network (PNN), generalized regression neural network, and support vector machine (SVM) are used as classifiers. A high success rate of 86.01% is obtained using SVM for the jackknife test. In case of independent dataset test, PNN yields the highest accuracy of 95.73%. These classifiers exhibit improved performance using other performance measures such as sensitivity, specificity, Mathew's correlation coefficient, and F-measure. The experimental results show that the prediction performance of the proposed scheme for classifying membrane protein types is the best reported, so far. This performance improvement may largely be credited to the learning capabilities of neural networks and the composite feature extraction strategy, which exploits seven different properties of protein sequences. The proposed Mem-Predictor can be accessed at http://111.68.99.218/Mem-Predictor.     

Protein contact prediction using patterns of correlation

We describe a new method for using neural networks to predict residue contact pairs in a protein. The main inputs to the neural network are a set of 25 measures of correlated mutation between all pairs of residues in two "windows" of size 5 centered on the residues of interest. While the individual pair-wise correlations are a relatively weak predictor of contact, by training the network on windows of correlation the accuracy of prediction is significantly improved. The neural network is trained on a set of 100 proteins and then tested on a disjoint set of 1033 proteins of known structure. An average predictive accuracy of 21.7% is obtained taking the best L/2 predictions for each protein, where L is the sequence length. Taking the best L/10 predictions gives an average accuracy of 30.7%. The predictor is also tested on a set of 59 proteins from the CASP5 experiment. The accuracy is found to be relatively consistent across different sequence lengths, but to vary widely according to the secondary structure. Predictive accuracy is also found to improve by using multiple sequence alignments containing many sequences to calculate the correlations.     

PPRODO: prediction of protein domain boundaries using neural networks

Successful prediction of protein domain boundaries provides valuable information not only for the computational structure prediction of multidomain proteins but also for the experimental structure determination. Since protein sequences of multiple domains may contain much information regarding evolutionary processes such as gene-exon shuffling, this information can be detected by analyzing the position-specific scoring matrix (PSSM) generated by PSI-BLAST. We have presented a method, PPRODO (Prediction of PROtein DOmain boundaries) that predicts domain boundaries of proteins from sequence information by a neural network. The network is trained and tested using the values obtained from the PSSM generated by PSI-BLAST. A 10-fold cross-validation technique is performed to obtain the parameters of neural networks using a nonredundant set of 522 proteins containing 2 contiguous domains. PPRODO provides good and consistent results for the prediction of domain boundaries, with accuracy of about 66% using the +/-20 residue criterion. The PPRODO source code, as well as all data sets used in this work, are available from http://gene.kias.re.kr/ approximately jlee/pprodo/.     

Prediction of protein secondary structure content using amino acid composition and evolutionary information

Knowing protein structure and inferring its function from the structure are one of the main issues of computational structural biology, and often the first step is studying protein secondary structure. There have been many attempts to predict protein secondary structure contents. Previous attempts assumed that the content of protein secondary structure can be predicted successfully using the information on the amino acid composition of a protein. Recent methods achieved remarkable prediction accuracy by using the expanded composition information. The overall average error of the most successful method is 3.4%. Here, we demonstrate that even if we only use the simple amino acid composition information alone, it is possible to improve the prediction accuracy significantly if the evolutionary information is included. The idea is motivated by the observation that evolutionarily related proteins share the similar structure. After calculating the homolog-averaged amino acid composition of a protein, which can be easily obtained from the multiple sequence alignment by running PSI-BLAST, those 20 numbers are learned by a multiple linear regression, an artificial neural network and a support vector regression. The overall average error of method by a support vector regression is 3.3%. It is remarkable that we obtain the comparable accuracy without utilizing the expanded composition information such as pair-coupled amino acid composition. This work again demonstrates that the amino acid composition is a fundamental characteristic of a protein. It is anticipated that our novel idea can be applied to many areas of protein bioinformatics where the amino acid composition information is utilized, such as subcellular localization prediction, enzyme subclass prediction, domain boundary prediction, signal sequence prediction, and prediction of unfolded segment in a protein sequence, to name a few.     

MUFOLD‐SS: New deep inception‐inside‐inception networks for protein secondary structure prediction

Protein secondary structure prediction can provide important information for protein 3D structure prediction and protein functions. Deep learning offers a new opportunity to significantly improve prediction accuracy. In this article, a new deep neural network architecture, named the Deep inception‐inside‐inception (Deep3I) network, is proposed for protein secondary structure prediction and implemented as a software tool MUFOLD‐SS. The input to MUFOLD‐SS is a carefully designed feature matrix corresponding to the primary amino acid sequence of a protein, which consists of a rich set of information derived from individual amino acid, as well as the context of the protein sequence. Specifically, the feature matrix is a composition of physio‐chemical properties of amino acids, PSI‐BLAST profile, and HHBlits profile. MUFOLD‐SS is composed of a sequence of nested inception modules and maps the input matrix to either eight states or three states of secondary structures. The architecture of MUFOLD‐SS enables effective processing of local and global interactions between amino acids in making accurate prediction. In extensive experiments on multiple datasets, MUFOLD‐SS outperformed the best existing methods and other deep neural networks significantly. MUFold‐SS can be downloaded from http://dslsrv8.cs.missouri.edu/~cf797/MUFoldSS/download.html .     

Discrimination of outer membrane proteins using machine learning algorithms

Discriminating outer membrane proteins (OMPs) from other folding types of globular and membrane proteins is an important task both for identifying OMPs from genomic sequences and for the successful prediction of their secondary and tertiary structures. In this work, we have analyzed the performance of different methods, based on Bayes rules, logistic functions, neural networks, support vector machines, decision trees, etc. for discriminating OMPs. We found that most of the machine learning techniques discriminate OMPs with similar accuracy. The neural network-based method could discriminate the OMPs from other proteins [globular/transmembrane helical (TMH)] at the fivefold cross-validation accuracy of 91.0% in a dataset of 1,088 proteins. The accuracy of discriminating globular proteins is 88.8% and that of TMH proteins is 93.7%. Further, the neural network method is tested with globular proteins belonging to 30 different folding types and it could successfully exclude 95% of the considered proteins. The proteins with SAM domain such as knottins, rubredoxin, and thioredoxin folds are eliminated with 100% accuracy. These accuracy levels are comparable to or better than other methods in the literature. We suggest that this method could be effectively used to discriminate OMPs and for detecting OMPs in genomic sequences.     

Prediction of protein folding class from amino acid composition

An empirical relation between the amino acid composition and three-dimensional folding pattern of several classes of proteins has been determined. Computer simulated neural networks have been used to assign proteins to one of the following classes based on their amino acid composition and size: (1) 4α-helical bundles, (2) parallel (α/β)₈ barrels, (3) nucleotide binding fold, (4) immunoglobulin fold, or (5) none of these. Networks trained on the known crystal structures as well as sequences of closely related proteins are shown to correctly predict folding classes of proteins not represented in the training set with an average accuracy of 87%. Other folding motifs can easily be added to the prediction scheme once larger databases become available. Analysis of the neural network weights reveals that amino acids favoring prediction of a folding class are usually over represented in that class and amino acids with unfavorable weights are underrepresented in composition. The neural networks utilize combinations of these multiple small variations in amino acid composition in order to make a prediction. The favorably weighted amino acids in a given class also form the most intramolecular interactions with other residues in proteins of that class. A detailed examination of the contacts of these amino acids reveals some general patterns that may help stabilize each folding class. © 1993 Wiley-Liss, Inc.     

An hierarchical artificial neural network system for the classification of transmembrane proteins.

This work presents a simple artificial neural network which classifies proteins into two classes from their sequences alone: the membrane protein class and the non-membrane protein class. This may be important in the functional assignment and analysis of open reading frames (ORF's) identified in complete genomes and, especially, those ORF's that correspond to proteins with unknown function. The network described here has a simple hierarchical feed-forward topology and a limited number of neurons which makes it very fast. By using only information contained in 11 protein sequences, the method was able to identify, with 100% accuracy, all membrane proteins with reliable topologies collected from several papers in the literature. Applied to a test set of 995 globular, water-soluble proteins, the neural network classified falsely 23 of them in the membrane protein class (97.7% of correct assignment). The method was also applied to the complete SWISS-PROT database with considerable success and on ORF's of several complete genomes. The neural network developed was associated with the PRED-TMR algorithm (Pasquier,C., Promponas,V.J., Palaios,G.A., Hamodrakas,J.S. and Hamodrakas,S.J., 1999) in a new application package called PRED-TMR2. A WWW server running the PRED-TMR2 software is available at http://o2.db.uoa.gr/PRED-TMR2     

Novel hybrid method for the evaluation of parameters contributing in determination of protein structural classes

Due to the increasing gap between structure-determined and sequenced proteins, prediction of protein structural classes has been an important problem. It is very important to use efficient sequential parameters for developing class predictors because of the close sequence-structure relationship. The multinomial logistic regression model was used for the first time to evaluate the contribution of sequence parameters in determining the protein structural class. An in-house program generated parameters including single amino acid and all dipeptide composition frequencies. Then, the most effective parameters were selected by a multinomial logistic regression. Selected variables in the multinomial logistic model were Valine among single amino acid composition frequencies and Ala-Gly, Cys-Arg, Asp-Cys, Glu-Tyr, Gly-Glu, His-Tyr, Lys-Lys, Leu-Asp, Leu-Arg, Pro-Cys, Gln-Met, Gln-Thr, Ser-Trp, Val-Asn and Trp-Asn among dipeptide composition frequencies. Also a neural network model was constructed and fed by the parameters selected by multinomial logistic regression to build a hybrid predictor. In this study, self-consistency and jackknife tests on a database constructed by Zhou [1998. An intriguing controversy over protein structural class prediction. J. Protein Chem. 17(8), 729-738] containing 498 proteins are used to verify the performance of this hybrid method, and are compared with some of prior works. The results showed that our two-stage hybrid model approach is very promising and may play a complementary role to the existing powerful approaches.     

Transmembrane helices predicted at 95% accuracy.

We describe a neural network system that predicts the locations of transmembrane helices in integral membrane proteins. By using evolutionary information as input to the network system, the method significantly improved on a previously published neural network prediction method that had been based on single sequence information. The input data were derived from multiple alignments for each position in a window of 13 adjacent residues: amino acid frequency, conservation weights, number of insertions and deletions, and position of the window with respect to the ends of the protein chain. Additional input was the amino acid composition and length of the whole protein. A rigorous cross-validation test on 69 proteins with experimentally determined locations of transmembrane segments yielded an overall two-state per-residue accuracy of 95%. About 94% of all segments were predicted correctly. When applied to known globular proteins as a negative control, the network system incorrectly predicted fewer than 5% of globular proteins as having transmembrane helices. The method was applied to all 269 open reading frames from the complete yeast VIII chromosome. For 59 of these, at least two transmembrane helices were predicted. Thus, the prediction is that about one-fourth of all proteins from yeast VIII contain one transmembrane helix, and some 20%, more than one.     

Prediction of alpha-turns in proteins using PSI-BLAST profiles and secondary structure information

In this paper a systematic attempt has been made to develop a better method for predicting alpha-turns in proteins. Most of the commonly used approaches in the field of protein structure prediction have been tried in this study, which includes statistical approach "Sequence Coupled Model" and machine learning approaches; i) artificial neural network (ANN); ii) Weka (Waikato Environment for Knowledge Analysis) Classifiers and iii) Parallel Exemplar Based Learning (PEBLS). We have also used multiple sequence alignment obtained from PSIBLAST and secondary structure information predicted by PSIPRED. The training and testing of all methods has been performed on a data set of 193 non-homologous protein X-ray structures using five-fold cross-validation. It has been observed that ANN with multiple sequence alignment and predicted secondary structure information outperforms other methods. Based on our observations we have developed an ANN-based method for predicting alpha-turns in proteins. The main components of the method are two feed-forward back-propagation networks with a single hidden layer. The first sequence-structure network is trained with the multiple sequence alignment in the form of PSI-BLAST-generated position specific scoring matrices. The initial predictions obtained from the first network and PSIPRED predicted secondary structure are used as input to the second structure-structure network to refine the predictions obtained from the first net. The final network yields an overall prediction accuracy of 78.0% and MCC of 0.16. A web server AlphaPred (http://www.imtech.res.in/raghava/alphapred/) has been developed based on this approach.     

XANNpred: neural nets that predict the propensity of a protein to yield diffraction-quality crystals

Production of diffracting crystals is a critical step in determining the three-dimensional structure of a protein by X-ray crystallography. Computational techniques to rank proteins by their propensity to yield diffraction-quality crystals can improve efficiency in obtaining structural data by guiding both protein selection and construct design. XANNpred comprises a pair of artificial neural networks that each predict the propensity of a selected protein sequence to produce diffraction-quality crystals by current structural biology techniques. Blind tests show XANNpred has accuracy and Matthews correlation values ranging from 75% to 81% and 0.50 to 0.63 respectively; values of area under the receiver operator characteristic (ROC) curve range from 0.81 to 0.88. On blind test data XANNpred outperforms the other available algorithms XtalPred, PXS, OB-Score, and ParCrys. XANNpred also guides construct design by presenting graphs of predicted propensity for diffraction-quality crystals against residue sequence position. The XANNpred-SG algorithm is likely to be most useful to target selection in structural genomics consortia, while the XANNpred-PDB algorithm is more suited to the general structural biology community. XANNpred predictions that include sliding window graphs are freely available from http://www.compbio.dundee.ac.uk/xannpred     

Achieving 80% ten-fold cross-validated accuracy for secondary structure prediction by large-scale training

An integrated system of neural networks, called SPINE, is established and optimized for predicting structural properties of proteins. SPINE is applied to three-state secondary-structure and residue-solvent-accessibility (RSA) prediction in this paper. The integrated neural networks are carefully trained with a large dataset of 2640 chains, sequence profiles generated from multiple sequence alignment, representative amino acid properties, a slow learning rate, overfitting protection, and an optimized sliding-widow size. More than 200,000 weights in SPINE are optimized by maximizing the accuracy measured by Q(3) (the percentage of correctly classified residues). SPINE yields a 10-fold cross-validated accuracy of 79.5% (80.0% for chains of length between 50 and 300) in secondary-structure prediction after one-month (CPU time) training on 22 processors. An accuracy of 87.5% is achieved for exposed residues (RSA >95%). The latter approaches the theoretical upper limit of 88-90% accuracy in assigning secondary structures. An accuracy of 73% for three-state solvent-accessibility prediction (25%/75% cutoff) and 79.3% for two-state prediction (25% cutoff) is also obtained.     

A deep dense inception network for protein beta-turn prediction

Beta-turn prediction is useful in protein function studies and experimental design. Although recent approaches using machine-learning techniques such as support vector machine (SVM), neural networks, and K nearest neighbor have achieved good results for beta-turn prediction, there is still significant room for improvement. As previous predictors utilized features in a sliding window of 4-20 residues to capture interactions among sequentially neighboring residues, such feature engineering may result in incomplete or biased features and neglect interactions among long-range residues. Deep neural networks provide a new opportunity to address these issues. Here, we proposed a deep dense inception network (DeepDIN) for beta-turn prediction, which takes advantage of the state-of-the-art deep neural network design of dense networks and inception networks. A test on a recent BT6376 benchmark data set shows that DeepDIN outperformed the previous best tool BetaTPred3 significantly in both the overall prediction accuracy and the nine-type beta-turn classification accuracy. A tool, called MUFold-BetaTurn, was developed, which is the first beta-turn prediction tool utilizing deep neural networks. The tool can be downloaded at http://dslsrv8.cs.missouri.edu/~cf797/MUFoldBetaTurn/download.html .     

Use of different approaches to model presence/absence of Salmo marmoratus in Piedmont (Northwestern Italy)

In Piedmont (Italy) the environmental changes due to human impact have had profound effects on rivers and their inhabitants. Thus, it is necessary to develop practical tools providing accurate ecological assessments of river and species conditions. We focus our attention on Salmo marmoratus, an endangered salmonid which is characteristic of the Po river system in Italy. In order to contribute to the management of the species, four different approaches were used to assess its presence: discriminant function analysis, logistic regression, decision tree models and artificial neural networks. Either all the 20 environmental variables measured in the field or the 7 coming from feature selection were used to classify sites as positive or negative for S. marmoratus. The performances of the different models were compared. Discriminant function analysis, logistic regression, and decision tree models (unpruned and pruned) had relatively high percentages of correctly classified instances. Although neither tree-pruning technique improved the reliability of the models significantly, they did reduce the tree complexity and hence increased the clarity of the models. The artificial neural network (ANN) approach, especially the model built with the 7 inputs coming from feature selection, showed better performance than all the others. The relative contribution of each independent variable to this model was determined by using the sensitivity analysis technique. Our findings proved that the ANNs were more effective than the other classification techniques. Moreover, ANNs achieved their high potentials when they were applied in models used to make decisions regarding river and conservation management.     

How many potentially secreted proteins are contained in a bacterial genome?

Artificial neural networks were trained on the prediction of the subcellular location of bacterial proteins. A cross-validated average prediction accuracy of 93% was reached for distinction between cytoplasmic and non-cytoplasmic proteins, based on the analysis of protein amino-acid composition. Principal component analysis and self-organizing maps were used to create graphical representations of amino-acid sequence space. A clear separation of cytoplasmic, periplasmic, and extracellular proteins was observed. The neural network system was applied to predicting potentially secreted proteins in 15 complete genomes. For mesophile bacteria the predicted fractions of non-cytoplasmic proteins agree with previously published estimates, ranging between 15% and 30%. Characteristics of thermophile genomes might lead to an under-estimation of the fraction of secreted proteins by presently available prediction systems. A self-organizing map was constructed from all 15 bacterial genomes. This technique can reveal additional sequence features independent from exhaustive pair-wise sequence alignment. The Treponema pallidum and Mycobacterium tuberculosis data formed separate clusters indicating unusual characteristics of these genomes.