Context-free and context-sensitive dynamics in recurrent neural networks

Continuous-valued recurrent neural networks can learn mechanisms for processing context-free languages. The dynamics of such networks is usually based on damped oscillation around fixed points in state space and requires that the dynamical components are arranged in certain ways. It is shown that qualitatively similar dynamics with similar constraints hold for a(n)b(n)c(n), a context-sensitive language. The additional difficulty with a(n)b(n)c(n), compared with the context-free language a(n)b(n), consists of 'counting up' and 'counting down' letters simultaneously. The network solution is to oscillate in two principal dimensions, one for counting up and one for counting down. This study focuses on the dynamics employed by the sequential cascaded network, in contrast to the simple recurrent network, and the use of backpropagation through time. Found solutions generalize well beyond training data, however, learning is not reliable. The contribution of this study lies in demonstrating how the dynamics in recurrent neural networks that process context-free languages can also be employed in processing some context-sensitive languages (traditionally thought of as requiring additional computation resources). This continuity of mechanism between language classes contributes to our understanding of neural networks in modelling language learning and processing.

An adaptive neural-wavelet model for short term load forecasting

This paper proposed a novel model for short term load forecast in the competitive electricity market. The prior electricity demand data are treated as time series. The forecast model is based on wavelet multi-resolution decomposition by autocorrelation shell representation and neural networks (multilayer perceptrons, or MLPs) modeling of wavelet coefficients. To minimize the influence of noisy low level coefficients, we applied the practical Bayesian method Automatic Relevance Determination (ARD) model to choose the size of MLPs, which are then trained to provide forecasts. The individual wavelet domain forecasts are recombined to form the accurate overall forecast. The proposed method is tested using Queensland electricity demand data from the Australian National Electricity Market. (C) 2001 Elsevier Science B.V. All rights reserved.

Probabilistic modelling of CO2 corrosion laboratory data using neural networks

The present paper addresses two major concerns that were identified when developing neural network based prediction models and which can limit their wider applicability in the industry. The first problem is that it appears neural network models are not readily available to a corrosion engineer. Therefore the first part of this paper describes a neural network model of CO2 corrosion which was created using a standard commercial software package and simple modelling strategies. It was found that such a model was able to capture practically all of the trends noticed in the experimental data with acceptable accuracy. This exercise has proven that a corrosion engineer could readily develop a neural network model such as the one described below for any problem at hand, given that sufficient experimental data exist. This applies even in the cases when the understanding of the underlying processes is poor. The second problem arises from cases when all the required inputs for a model are not known or can be estimated with a limited degree of accuracy. It seems advantageous to have models that can take as input a range rather than a single value. One such model, based on the so-called Monte Carlo approach, is presented. A number of comparisons are shown which have illustrated how a corrosion engineer might use this approach to rapidly test the sensitivity of a model to the uncertainities associated with the input parameters. (C) 2001 Elsevier Science Ltd. All rights reserved.

On learning context-free and context-sensitive languages

The long short-term memory (LSTM) is not the only neural network which learns a context sensitive language. Second-order sequential cascaded networks (SCNs) are able to induce means from a finite fragment of a context-sensitive language for processing strings outside the training set. The dynamical behavior of the SCN is qualitatively distinct from that observed in LSTM networks. Differences in performance and dynamics are discussed.

The parameterless self-organizing map algorithm

The parameterless self-organizing map (PLSOM) is a new neural network algorithm based on the self-organizing map (SOM). It eliminates the need for a learning rate and annealing schemes for learning rate and neighborhood size. We discuss the relative performance of the PLSOM and the SOM and demonstrate some tasks in which the SOM fails but the PLSOM performs satisfactory. Finally we discuss some example applications of the PLSOM and present a proof of ordering under certain limited conditions.

Tracking the states of a nonlinear and nonstationary system in the weight-space of artificial neural networks

We propose a novel interpretation and usage of Neural Network (NN) in modeling physiological signals, which are allowed to be nonlinear and/or nonstationary. The method consists of training a NN for the k-step prediction of a physiological signal, and then examining the connection-weight-space (CWS) of the NN to extract information about the signal generator mechanism. We de. ne a novel feature, Normalized Vector Separation (gamma(ij)), to measure the separation of two arbitrary states i and j in the CWS and use it to track the state changes of the generating system. The performance of the method is examined via synthetic signals and clinical EEG. Synthetic data indicates that gamma(ij) can track the system down to a SNR of 3.5 dB. Clinical data obtained from three patients undergoing carotid endarterectomy of the brain showed that EEG could be modeled (within a root-means-squared-error of 0.01) by the proposed method, and the blood perfusion state of the brain could be monitored via gamma(ij), with small NNs having no more than 21 connection weight altogether.

Detecting and sorting targeting peptides wtih neural networks and support vector machines

This paper presents a composite multi-layer classifier system for predicting the subcellular localization of proteins based on their amino acid sequence. The work is an extension of our previous predictor PProwler v1.1 which is itself built upon the series of predictors SignalP and TargetP. In this study we outline experiments conducted to improve the classifier design. The major improvement came from using Support Vector machines as a "smart gate" sorting the outputs of several different targeting peptide detection networks. Our final model (PProwler v1.2) gives MCC values of 0.873 for non-plant and 0.849 for plant proteins. The model improves upon the accuracy of our previous subcellular localization predictor (PProwler v1.1) by 2% for plant data (which represents 7.5% improvement upon TargetP).

Strategies for identifying and predicting islet autoantigen T-cell epitopes in insulin-dependent diabetes mellitus

T cells recognize peptide epitopes bound to major histocompatibility complex molecules. Human T-cell epitopes have diagnostic and therapeutic applications in autoimmune diseases. However, their accurate definition within an autoantigen by T-cell bioassay, usually proliferation, involves many costly peptides and a large amount of blood, We have therefore developed a strategy to predict T-cell epitopes and applied it to tyrosine phosphatase IA-2, an autoantigen in IDDM, and HLA-DR4(*0401). First, the binding of synthetic overlapping peptides encompassing IA-2 was measured directly to purified DR4. Secondly, a large amount of HLA-DR4 binding data were analysed by alignment using a genetic algorithm and were used to train an artificial neural network to predict the affinity of binding. This bioinformatic prediction method was then validated experimentally and used to predict DR4 binding peptides in IA-2. The binding set encompassed 85% of experimentally determined T-cell epitopes. Both the experimental and bioinformatic methods had high negative predictive values, 92% and 95%, indicating that this strategy of combining experimental results with computer modelling should lead to a significant reduction in the amount of blood and the number of peptides required to define T-cell epitopes in humans.

Relationship between peptide selectivities of human transporters associated with antigen processing and HLA class I molecules

Efficiency of presentation of a peptide epitope by a MHC class I molecule depends on two parameters: its binding to the MHC molecule and its generation by intracellular Ag processing. In contrast to the former parameter, the mechanisms underlying peptide selection in Ag processing are poorly understood. Peptide translocation by the TAP transporter is required for presentation of most epitopes and may modulate peptide supply to MHC class I molecules. To study the role of human TAP for peptide presentation by individual HLA class I molecules, we generated artificial neural networks capable of predicting the affinity of TAP for random sequence 9-mer peptides. Using neural network-based predictions of TAP affinity, we found that peptides eluted from three different HLA class I molecules had higher TAP affinities than control peptides with equal binding affinities for the same HLA class I molecules, suggesting that human TAP may contribute to epitope selection. In simulated TAP binding experiments with 408 HLA class I binding peptides, HLA class I molecules differed significantly with respect to TAP affinities of their ligands, As a result, some class I molecules, especially HLA-B27, may be particularly efficient in presentation of cytosolic peptides with low concentrations, while most class I molecules may predominantly present abundant cytosolic peptides.

Control of fed-batch fermentations

Fed-batch fermentation is used to prevent or reduce substrate-associated growth inhibition by controlling nutrient supply. Here we review the advances in control of fed-batch fermentations. Simple exponential feeding and inferential methods are examined, as are newer methods based on fuzzy control and neural networks. Considerable interest has developed in these more advanced methods that hold promise for optimizing fed-batch techniques for complex fermentation systems. (C) 1999 Elsevier Science Inc. All rights reserved.

Computational binding assays of antigenic peptides

Computer models can be combined with laboratory experiments for the efficient determination of (i) peptides that bind MHC molecules and (ii) T-cell epitopes. For maximum benefit, the use of computer models must be treated as experiments analogous to standard laboratory procedures. This requires the definition of standards and experimental protocols for model application. We describe the requirements for validation and assessment of computer models. The utility of combining accurate predictions with a limited number of laboratory experiments is illustrated by practical examples. These include the identification of T-cell epitopes from IDDM-, melanoma- and malaria-related antigens by combining computational and conventional laboratory assays. The success rate in determining antigenic peptides, each in the context of a specific HLA molecule, ranged from 27 to 71%, while the natural prevalence of MHC-binding peptides is 0.1-5%.

Knowledge discovery and data mining in biological databases

The new technologies for Knowledge Discovery from Databases (KDD) and data mining promise to bring new insights into a voluminous growing amount of biological data. KDD technology is complementary to laboratory experimentation and helps speed up biological research. This article contains an introduction to KDD, a review of data mining tools, and their biological applications. We discuss the domain concepts related to biological data and databases, as well as current KDD and data mining developments in biology.

Efficient discovery of immune response targets by cyclical refinement of QSAR models of peptide binding

Peptides that induce and recall T-cell responses are called T-cell epitopes. T-cell epitopes may be useful in a subunit vaccine against malaria. Computer models that simulate peptide binding to MHC are useful for selecting candidate T-cell epitopes since they minimize the number of experiments required for their identification. We applied a combination of computational and immunological strategies to select candidate T-cell epitopes. A total of 86 experimental binding assays were performed in three rounds of identification of HLA-All binding peptides from the six preerythrocytic malaria antigens. Thirty-six peptides were experimentally confirmed as binders. We show that the cyclical refinement of the ANN models results in a significant improvement of the efficiency of identifying potential T-cell epitopes. (C) 2001 by Elsevier Science Inc.

Methods for prediction of peptide binding to MHC molecules: A comparative study

Background: A variety of methods for prediction of peptide binding to major histocompatibility complex (MHC) have been proposed. These methods are based on binding motifs, binding matrices, hidden Markov models (HMM), or artificial neural networks (ANN). There has been little prior work on the comparative analysis of these methods. Materials and Methods: We performed a comparison of the performance of six methods applied to the prediction of two human MHC class I molecules, including binding matrices and motifs, ANNs, and HMMs. Results: The selection of the optimal prediction method depends on the amount of available data (the number of peptides of known binding affinity to the MHC molecule of interest), the biases in the data set and the intended purpose of the prediction (screening of a single protein versus mass screening). When little or no peptide data are available, binding motifs are the most useful alternative to random guessing or use of a complete overlapping set of peptides for selection of candidate binders. As the number of known peptide binders increases, binding matrices and HMM become more useful predictors. ANN and HMM are the predictive methods of choice for MHC alleles with more than 100 known binding peptides. Conclusion: The ability of bioinformatic methods to reliably predict MHC binding peptides, and thereby potential T-cell epitopes, has major implications for clinical immunology, particularly in the area of vaccine design.