Analysis of protein transmembrane helical regions by a neural network.

Neural networks were used to generalize common themes found in transmembrane-spanning protein helices. Various-sized databases were used containing nonoverlapping sequences, each 25 amino acids long. Training consisted of sorting these sequences into 1 of 2 groups: transmembrane helical peptides or nontransmembrane peptides. Learning was measured using a test set 10% the size of the training set. As training set size increased from 214 sequences to 1,751 sequences, learning increased in a nonlinear manner from 75% to a high of 98%, then declined to a low of 87%. The final training database consisted of roughly equal numbers of transmembrane (928) and nontransmembrane (1,018) sequences. All transmembrane sequences were entered into the database with respect to their lipid membrane orientation: from inside the membrane to outside. Generalized transmembrane helix and nontransmembrane peptides were constructed from the maximally weighted connecting strengths of fully trained networks. Four generalized transmembrane helices were found to contain 9 consensus residues: a K-R-F triplet was found at the inside lipid interface, 2 isoleucine and 2 other phenylalanine residues were present in the helical body, and 2 tryptophan residues were found near the outside lipid interface. As a test of the training method, bacteriorhodopsin was examined to determine the position of its 7 transmembrane helices.     

Predicting protein secondary structure content. A tandem neural network approach.

A priori knowledge of secondary structure content can be of great use in theoretical and experimental determination of protein structure. We present a method that uses two computer-simulated neural networks placed in "tandem" to predict the secondary structure content of water-soluble, globular proteins. The first of the two networks, NET1, predicts a protein's helix and strand content given information about the protein's amino acid composition, molecular weight and heme presence. Because NET1 contained more adjustable parameters (network weights) than learning examples, this network experienced problems with memorization, which is the inability to generalize onto new, never-seen-before examples. To overcome this problem, we designed a second network, NET2, which learned to determine when NET1 was in a state of generalization. Together, these two networks produce prediction errors as low as 5.0% and 5.6% for helix and strand content, respectively, on a set of protein crystal structures bearing little homology to those used in network training. A comparison between three other methods including a multiple linear regression analysis, a non-hidden-node network analysis and a secondary structure assignment analysis reveals that our tandem neural network scheme is, indeed, the best method for predicting secondary structure content. The results of our analysis suggest that the knowledge of sequence information is not necessary for highly accurate predictions of protein secondary structure content.     

Evolutionary optimization and neural network models of behavior

One of the main challenges to the adaptionist program in general and the use of optimization models in behavioral and evolutionary ecology, in particular, is that organisms are so constrained' by ontogeny and phylogeny that they may not be able to attain optimal solutions, however those are defined. This paper responds to the challenge through the comparison of optimality and neural network models for the behavior of an individual polychaete worm. The evolutionary optimization model is used to compute behaviors (movement in and out of a tube) that maximize a measure of Darwinian fitness based on individual survival and reproduction. The neural network involves motor, sensory, energetic reserve and clock neuronal groups. Ontogeny of the neural network is the change of connections of a single individual in response to its experiences in the environment. Evolution of the neural network is the natural selection of initial values of connections between groups and learning rules for changing connections. Taken together, these can be viewed as design parameters. The best neural networks have fitnesses between 85% and 99% of the fitness of the evolutionary optimization model. More complicated models for polychaete worms are discussed. Formulation of a neural network model for host acceptance decisions by tephritid fruit flies leads to predictions about the neurobiology of the flies. The general conclusion is that neural networks appear to be sufficiently rich and plastic that even weak evolution of design parameters may be sufficient for organisms to achieve behaviors that give fitnesses close to the evolutionary optimal fitness, particularly if the behaviors are relatively simple.     

Identification of Cryptosporidium parvum oocysts by an artificial neural network approach

Microscopic detection of Cryptosporidium parvum oocysts is time-consuming, requires trained analysts, and is frequently subject to significant human errors. Artificial neural networks (ANN) were developed to help identify immunofluorescently labeled C. parvum oocysts. A total of 525 digitized images of immunofluorescently labeled oocysts, fluorescent microspheres, and other miscellaneous nonoocyst images were employed in the training of the ANN. The images were cropped to a 36- by 36-pixel image, and the cropped images were placed into two categories, oocyst and nonoocyst images. The images were converted to grayscale and processed into a histogram of gray color pixel intensity. Commercially available software was used to develop and train the ANN. The networks were optimized by varying the number of training images, number of hidden neurons, and a combination of these two parameters. The network performance was then evaluated using a set of 362 unique testing images which the network had never "seen" before. Under optimized conditions, the correct identification of authentic oocyst images ranged from 81 to 97%, and the correct identification of nonoocyst images ranged from 78 to 82%, depending on the type of fluorescent antibody that was employed. The results indicate that the ANN developed were able to generalize the training images and subsequently discern previously unseen oocyst images efficiently and reproducibly. Thus, ANN can be used to reduce human errors associated with the microscopic detection of Cryptosporidium oocysts.     

Higherorder neural network group models for financial simulation

Real world financial data is often discontinuous and non-smooth. If we attempt to use neural networks to simulate such functions, then accuracy will be a problem. Neural network group models perform this function much better. Both Polynomial Higher Order Neural network Group (PHONG) and Trigonometric polynomial Higher Order Neural network Group (THONG) models are developed. These HONG models are open box, convergent models capable of approximating any kind of piecewise continuous function, to any degree of accuracy. Moreover they are capable of handling higher frequency, higher order non-linear and discontinuous data. Results obtained using a Higher Order Neural network Group financial simulator are presented, which confirm that HONG group models converge without difficulty, and are considerably more accurate than neural network models (more specifically, around twice as good for prediction, and a factor of four improvement in the case of simulation).     

The immune system as a neural network: a multi-epitope approach.

The term "neural network" has been applied to arrays of simple activation units linked by weighted connections. If the connections are modified according to a defined learning algorithm, such networks can be trained to store and retrieve patterned information. Memories are distributed throughout the network, allowing the network to recall complete patterns from incomplete input (pattern completion). The major biological application of neural network theory to date has been in the neurosciences, but the immune system may represent an alternative organ system in which to search for neural network architecture. Previous applications of parallel distributed processing to idiotype network theory have focused upon the recognition of individual epitopes. We argue here that this approach may be too restrictive, underestimating the power of neural network architecture. We propose that the network stores and retrieves large, complex patterns consisting of multiple epitopes separated in time and space. Such a network would be capable of perceiving an entire bacterium, and of storing the time course of a viral infection. While recognition of solitary epitopes occurs at the cellular level in this model, recognition of structures larger than the width of an antibody binding site takes place at the organ level, via network architecture integration of, i.e. individual epitope responses. The Oudin-Cazenave enigma, the sharing of idiotypic determinants by antibodies directed against distinct regions of the same antigen, suggests that some network level of integration of the individual clonal responses to large antigens does occur. The role of cytokines in prior neural network models of the immune system is unclear. We speculate that cytokines may influence the temperature of the network, such that changes in the cytokine milieu serve to "anneal" the network, allowing it to achieve the optimum steady-state in the shortest period of time.     

Prediction of protein secondary structure by an enhanced neural network.

Computational model of neural network is used for prediction of secondary structure of globular proteins of known sequence. In contrast to earlier works some information about expected tertiary interactions were built in into the neural network. As a result the prediction accuracy was improved by 3% to 5%. Possible applications of this new approach are briefly discussed.     

Mapping of protein surface cavities and prediction of enzyme class by a self-organizing neural network

An automated computer-based method for mapping of protein surface cavities was developed and applied to a set of 176 metalloproteinases containing zinc cations in their active sites. With very few exceptions, the cavity search routine detected the active site among the five largest cavities and produced reasonable active site surfaces. Cavities were described by means of solvent-accessible surface patches. For a given protein, these patches were calculated in three steps: (i) definition of cavity atoms forming surface cavities by a grid-based technique; (ii) generation of solvent accessible surfaces; (iii) assignment of an accessibility value and a generalized atom type to each surface point. Topological correlation vectors were generated from the set of surface points forming the cavities, and projected onto the plane by a self-organizing network. The resulting map of 865 enzyme cavities displays clusters of active sites that are clearly separated from the other cavities. It is demonstrated that both fully automated recognition of active sites, and prediction of enzyme class can be performed for novel protein structures at high accuracy.     

Optimization of ion-exchange protein separations using a vector quantizing neural network

In this work, a previously proposed methodology for the optimization of analytical scale protein separations using ion-exchange chromatography is subjected to two challenging case studies. The optimization methodology uses a Doehlert shell design for design of experiments and a novel criteria function to rank chromatograms in order of desirability. This chromatographic optimization function (COF) accounts for the separation between neighboring peaks, the total number of peaks eluted, and total analysis time. The COF is penalized when undesirable peak geometries (i.e., skewed and/or shouldered peaks) are present as determined by a vector quantizing neural network. Results of the COF analysis are fit to a quadratic response model, which is optimized with respect to the optimization variables using an advanced Nelder and Mead simplex algorithm. The optimization methodology is tested on two case study sample mixtures, the first of which is composed of equal parts of lysozyme, conalbumin, bovine serum albumin, and transferrin, and the second of which contains equal parts of conalbumin, bovine serum albumin, tranferrin, beta-lactoglobulin, insulin, and alpha -chymotrypsinogen A. Mobile-phase pH and gradient length are optimized to achieve baseline resolution of all solutes for both case studies in acceptably short analysis times, thus demonstrating the usefulness of the empirical optimization methodology.     

A maximum neural network approach for N-queens problems

 A novel neural network approach using the maximum neuron model is presented for N-queens problems. The goal of the N-queens problem is to find a set of locations of N queens on an N×N chessboard such that no pair of queens commands each other. The maximum neuron model proposed by Takefuji et al. has been applied to two optimization problems where the optimization of objective functions is requested without constraints. This paper demonstrates the effectiveness of the maximum neuron model for constraint satisfaction problems through the N-queens problem. The performance is verified through simulations in up to 500-queens problems on the sequential mode, the N-parallel mode, and the N ²-parallel mode, where our maximum neural network shows the far better performance than the existing neural networks. Received: 4 June 1996/Accepted in revised form: 13 November 1996     

Limits on alpha-helix prediction with neural network models.

Using a backpropagation neural network model we have found a limit for secondary structure prediction from local sequence. By including only sequences from whole alpha-helix and non-alpha-helix structures in our training and test sets--sequences spanning boundaries between these two structures were excluded--it was possible to investigate directly the relationship between sequence and structure for alpha-helix. A group of non-alpha-helix sequences, that was disrupting overall prediction success, was indistinguishable to the network from alpha-helix sequences. These sequences were found to occur at regions adjacent to the termini of alpha-helices with statistical significance, suggesting that potentially longer alpha-helices are disrupted by global constraints. Some of these regions spanned more than 20 residues. On these whole structure sequences, 10 residues in length, a comparatively high prediction success of 78% with a correlation coefficient of 0.52 was achieved. In addition, the structure of the input space, the distribution of beta-sheet in this space, and the effect of segment length were also investigated.     

Improvements in protein secondary structure prediction by an enhanced neural network

Computational neural networks have recently been used to predict the mapping between protein sequence and secondary structure. They have proven adequate for determining the first-order dependence between these two sets, but have, until now, been unable to garner higher-order information that helps determine secondary structure. By adding neural network units that detect periodicities in the input sequence, we have modestly increased the secondary structure prediction accuracy. The use of tertiary structural class causes a marked increase in accuracy. The best case prediction was 79% for the class of all-alpha proteins. A scheme for employing neural networks to validate and refine structural hypotheses is proposed. The operational difficulties of applying a learning algorithm to a dataset where sequence heterogeneity is under-represented and where local and global effects are inadequately partitioned are discussed.     

Prediction of siRNA knockdown efficiency using artificial neural network models

Selective knockdown of gene expression by short interference RNAs (siRNAs) has allowed rapid validation of gene functions and made possible a high throughput, genome scale approach to interrogate gene function. However, randomly designed siRNAs display different knockdown efficiencies of target genes. Hence, various prediction algorithms based on siRNA functionality have recently been constructed to increase the likelihood of selecting effective siRNAs, thereby reducing the experimental cost. Toward this end, we have trained three Back-propagation and Bayesian neural network models, previously not used in this context, to predict the knockdown efficiencies of 180 experimentally verified siRNAs on their corresponding target genes. Using our input coding based primarily on RNA structure thermodynamic parameters and cross-validation method, we showed that our neural network models outperformed most other methods and are comparable to the best predicting algorithm thus far published. Furthermore, our neural network models correctly classified 74% of all siRNAs into different efficiency categories; with a correlation coefficient of 0.43 and receiver operating characteristic curve score of 0.78, thus highlighting the potential utility of this method to complement other existing siRNA classification and prediction schemes.     

Experimental validation for quantitative protein network models

Cellular responses are the consequence of complex reactions of protein networks. The complexity should ultimately be described by a set of formulas in a quantitative fashion, in which each formula defines the reactions in response to given types of input. However, testing these formulas has not been a simple task because of the lack of appropriate means for experimental validation. 'Reverse-phase' lysate microarrays have been proved to be powerful for such requirements and thus can be a good resource for providing an experimental reference point for the theoretical biology of protein networks.     

An artificial neural network approach for studying phytoplankton succession

Artificial neural networks are used to model phytoplankton succession and gain insight into the relative strengths of bottom-up and top-down forces shaping seasonal patterns in phytoplankton biomass and community composition. Model comparisons indicate that patterns in chlorophyll aconcentrations response instantaneously to patterns in nutrient concentrations (phosphorous (P), nitrite and nitrate (NO₂/NO₃–N) and ammonium (NH₄–H) concentrations) and zooplankton biomass (daphnid cladocera and copepoda biomass); whereas lagged responses in an index of algal community composition are evident. A randomization approach to neural networks is employed to reveal individual and interacting contributions of nutrient concentrations and zooplankton biomass to predictions of phytoplankton biomass and community composition. The results show that patterns in chlorophyll aconcentrations are directly associated with P, NO₂/NO₃–N and daphnid cladocera biomass, as well as related to interactions between daphnid cladocera biomass, and NO₂/NO₃–N and P. Similarly, patterns in phytoplankton community composition are associated with NO₂/NO₃–N and daphnid cladocera biomass; however show contrasting patterns in nutrient– zooplankton and zooplankton–zooplankton interactions. Together, the results provide correlative evidence for the importance of nutrient limitation, zooplankton grazing and nutrient regeneration in shaping phytoplankton community dynamics. This study shows that artificial neural networks can provide a powerful tool for studying phytoplankton succession by aiding in the quantification and interpretation of the individual and interacting contributions of nutrient limitation and zooplankton herbivory on phytoplankton biomass and community composition under natural conditions.     

Predicting the secondary structure of globular proteins using neural network models

We present a new method for predicting the secondary structure of globular proteins based on non-linear neural network models. Network models learn from existing protein structures how to predict the secondary structure of local sequences of amino acids. The average success rate of our method on a testing set of proteins non-homologous with the corresponding training set was 64.3% on three types of secondary structure (alpha-helix, beta-sheet, and coil), with correlation coefficients of C alpha = 0.41, C beta = 0.31 and Ccoil = 0.41. These quality indices are all higher than those of previous methods. The prediction accuracy for the first 25 residues of the N-terminal sequence was significantly better. We conclude from computational experiments on real and artificial structures that no method based solely on local information in the protein sequence is likely to produce significantly better results for non-homologous proteins. The performance of our method of homologous proteins is much better than for non-homologous proteins, but is not as good as simply assuming that homologous sequences have identical structures.     

Fitting of dynamic recurrent neural network models to sensory stimulus-response data

We present a theoretical study aiming at model fitting for sensory neurons. Conventional neural network training approaches are not applicable to this problem due to lack of continuous data. Although the stimulus can be considered as a smooth time-dependent variable, the associated response will be a set of neural spike timings (roughly the instants of successive action potential peaks) that have no amplitude information. A recurrent neural network model can be fitted to such a stimulus-response data pair by using the maximum likelihood estimation method where the likelihood function is derived from Poisson statistics of neural spiking. The universal approximation feature of the recurrent dynamical neuron network models allows us to describe excitatory-inhibitory characteristics of an actual sensory neural network with any desired number of neurons. The stimulus data are generated by a phased cosine Fourier series having a fixed amplitude and frequency but a randomly shot phase. Various values of amplitude, stimulus component size, and sample size are applied in order to examine the effect of the stimulus to the identification process. Results are presented in tabular and graphical forms at the end of this text. In addition, to demonstrate the success of this research, a study involving the same model, nominal parameters and stimulus structure, and another study that works on different models are compared to that of this research.     

Visualization of the nature of protein folding by a study of a distance constraint approach in two-dimensional models

A distance constraint approach is applied to two-dimensional models of proteins in order to visualize the nature of protein folding and to examine the relative roles of different ranges of interaction. Three different native structures (I, II, and III) are considered; they have two different kinds of residues, viz., hydrophobic and hydrophilic, and different sequences of these residues. We examine how the distance constraint approach functions in the prediction of protein folding when we know the sequence of the residues, the (fixed) bond lengths, the mean distances between residues i and i + 2, and i and i + 3, and the mean distances for hydrophobic–hydrophobic, hydrophobic–hydrophilic, and hydrophilic–hydrophilic contacts between residues i and i + j, where j ≥ 4. This approach involves optimization of an object function with respect to 98 variables and is not free of the multiple-minimum problem. The optimization is always terminated if the chain is entangled and/or the segments (residues) are packed too compactly to move. In order to escape from such situations and to take the excluded-volume effect into account, a Monte Carlo method is used after the optimization is trapped in local minima. Success in the prediction of folding is found to depend on the starting conformations and on the native conformations. Fair success is obtained in predicting the helix-like structure in protein I and the overall structure of protein III, but not the β-like structures of proteins I and II. Insofar as the prediction of the structure of protein III is reasonable, it appears that some sequences of residues produce greater constraints on their conformations than others, if one considers only the hydrophobic and hydrophilic nature of the residues. These results imply that, in the folding of real proteins in three dimensions, the competition for hydrophobic (and hydrophilic) residues for inside (outside) positions in the molecule probably constitutes a necessary but not a sufficient condition to form and stabilize the native structure. The failure to predict the structure of protein II, and part of that of protein I, suggests that there are two types of long-range interactions. One (which we considered here) is nonspecific (i.e., is defined only in terms of contacts between residues of the same or different polarity) and acts at any stage of protein folding; the other (which we did not consider here) is a specific interaction between residues in pairs and contributes only when the residues in the specific pair take on the native conformation. Presumably, incorporation of such specific long-range interactions, together with the nonspecific ones, is necessary for successful protein folding, using the distance constraint approach.