Biological Applications of Artificial Neural Networks
P. Baldi

Artificial neural networks have had a number of successfull applications in biology. Most of these applications are actually unrelated to biological neural networks, but rather focus on pattern recognition and data mining problems within the large data sets generated by new data aquisition technologies, such as genome sequencing or DNA gene expression microarrays. In this tutorial, we will first review neural networks within the more general Bayesian probabilistic framework for data analysis. We will then cover a number of applications such as: (a) pattern recognition in DNA and proteins; (b) protein structure prediction; (c) analysis and clustering of gene expression data; (d) modeling gene networks.

Protein ß-Sheet Partner Prediction by Neural Networks
P. Baldi, G. Pollastri, C. A. F. Andersen, S Brunak

Predicting the secondary structure (alpha-helices, ß-sheets, coils) of proteins is an important step towards understanding their three dimensional conformations. Unlike alpha-helices that are built up from one contiguous region of the polypeptide chain, ß-sheets are more complex resulting from a combination several disjoint regions. The exact nature of these long distance interactions remains unclear. Here we introduce a neural-network based method for the prediction of amino acid partners in parallel as well as anti-parallel ß-sheets. The neural architecture predicts whether two residues located at the center of two distant windows are paired or not in a ß-sheet structure. The distance between the windows is a third essential input into the architecture. Variations on this architecture are trained using a large corpus of curated data. Prediction on both coupled and non-coupled residues currently exceeds 83% accuracy, well above any previously reported method. Unlike standard secondary structure prediction methods, the use of multiple alignment (profiles) in our case seems to degrade the performance, probably as a result of intra-chain correlation effects.

ART Neural Networks for Medical Data Analysis and Fast Distributed Learning
G. A. Carpenter, B. L. Milenova

ART (Adaptive Resonance Theory) neural networks for fast, stable learning and prediction have been applied in a variety of areas. Applications include airplane design and manufacturing, automatic target recognition, financial forecasting, machine tool monitoring, digital circuit design, chemical analysis, and robot vision. Supervised ART architectures, called ARTMAP systems, feature internal control mechanisms that create stable recognition categories of optimal size by maximizing code compression while minimizing predictive error in an on-line setting. Special-purpose requirements of various application domains have led to a number of ARTMAP variants, including fuzzy ARTMAP, ART-EMAP, Gaussian ARTMAP, and distributed ARTMAP. The talk at the ANNIMAB-1 conference (Göteborg, Sweden, May, 2000) will outline some ARTMAP applications, including computer-assisted medical diagnosis. Medical databases present many of the challenges found in general information management settings where speed, efficiency, ease of use, and accuracy are at a premium. A direct goal of improved computer-assisted medicine is to help deliver quality emergency care in situations that may be less than ideal. Working with these problems has stimulated a number of ART architecture developments, including ARTMAP-IC. A recent collaborative effort, using a new cardiac care database for system development, has brought together medical statisticians and clinicians at the New England Medical Center with researchers developing expert systems and neural networks, in order to create a hybrid method for medical diagnosis. The talk will also consider new neural network architectures, including distributed ART (dART), a real-time model of parallel distributed pattern learning that permits fast as well as slow adaptation, without catastrophic forgetting. Local synaptic computations in the dART model quantitatively match the paradoxical phenomenon of Markram-Tsodyks redistribution of synaptic efficacy, as a consequence of global system hypotheses.

Modelling Uncertainty in Biomedical Applications of Neural Networks
G. Dorffner, P. Sykacek, C. C. Schittenkopf

In this paper we argue that the explicit account of uncertainty in data modeling is particularly important for biomedical applications of neural networks and related techniques. There are several sources of uncertainty of a model, including noise, bias and variance. Unless one attempts to identify or minimize the sources that contribute to errors of a particular application, one only has a sub-optimal solution. If, on the other hand, one does attempt to model uncertainty, one gets several major advantages. We discuss several methods for modeling uncertainty, including density estimation, Bayesian inference and complex noise models, in the context of several sample applications - most notably in the domain of biosignal processing.

Neural Computation in Medicine: Perspectives and Prospects
R. Dybowski

In 1998, over 400 papers on artificial neural networks (ANNs) were published in the context of medicine, but why is there this interest in ANNs? And how do ANNs compare with traditional statistical methods? We propose some answers to these questions, and go on to consider the 'black box' issue.Finally,we briefly look at two directions in which ANNs are likely to develop, namely the use of Bayesian statistics and knowledge-data fusion.

Discriminating Gourmets, Lovers and Enophiles? Neural Nets Tell All About Locusts, Toads, and Roaches
W. M. Getz, W. C. Lemon

Here we consider the issue of choice and how neural systems can be used to investigate the processes of discrimination, as well as the evolution of different kinds of choice-related behavior in animals. We develop these ideas in the context of three studies, among others. The first study is on the evolution of specialization in animals using locust feeding behavior as the leitmotif, where decision making in individuals is modeled by a 3-layer-perceptron. In this study the fitness of individuals depends on their response to signals from plants and the density of individuals using those plants. The second is a study that investigates the evolution of species recognition in sympatric taxa using female mate choice in frogs as the leitmotif. Here individuals are modeled by Elman nets (3-layered perceptrons with feedback) and their fitness is determined by their ability to discriminate conspecifics from heterospecifics. The third is a study of the response characteristics of a recurrent Hopfield-type neural network to input that represents olfactory stimuli. The connectivity of this net reflects the basic architectural features of the neuron in the insect antennal lobe, as typified by cockroaches or bees.

An Unsupervised Learning Method that Produces Organized Representations from Real Information
T. Kohonen

The neural-network theories aim at two goals in medicine and biology: modeling of the neural structures and functions, and development of computational methods for the analysis of the experimental data. The Self-Organizing Map (SOM) was originally intended for the explanation of certain brain functions and organizations, but it has later been accepted as a new statistical analysis method to many fields of science and technology. At least 3700 scientific works on the SOM have been published. In its basic form, the SOM forms illustrative nonlinear projections of high-dimensional data manifolds, and these projections, usually produced on a two-dimensional display grid, help in the visualization and understanding of the relationships between complex data sets.

On Forgetful Attractor Network Memories
A. Lansner, A. Sandberg, K. M. Petersson, M. Ingvar

A recurrently connected attractor neural network with a Hebbian learning rule is currently our best ANN analogy for a piece of cortex. Functionally biological memory operates on a spectrum of time scales with regard to induction and retention, and it is modulated in complex ways by sub-cortical neuro-modulatory systems. Moreover, biological memory networks are commonly believed to be highly distributed and engage many co-operating cortical areas. Here we focus on the temporal aspects of induction and retention of memory in a connectionist type attractor memory model of a piece of cortex. A continuous time, forgetful Bayesian-Hebbian learning rule is described and compared to the characteristics of LTP and LTD seen experimentally. More generally, an attractor network implementing this learning rule can operate as a long-term, intermediate-term, or short-term memory. Modulation of the print-now signal of the learning rule replicates some experimental memory phenomena, like e.g. the von Restorff effect.

Outstanding Issues for Clinical Decision Support with Neural Networks
P. J. G. Lisboa, A. Vellido, H. Wong

Neural networks are widely used in potential medical applications, going as far as their introduction into commercial products. However, these pilot studies often ignore important aspects for clinical decision support. The context for the development of new medical technology is introduced, outstanding issues of principle for clinical support are identified, and their technical implications for neural network methodology are discussed.