Alistair G. Rust

Using Feature Selection Filtering Methods for Binding Site Predictions

Currently the best algorithms for transcription factor binding site prediction are severely limited in accuracy. In previous work we applied classification techniques on predictions from 12 key prediction algorithms. In this paper, we investigate the classification results when 4 feature selection filtering methods are used. They are bi-normal separation, correlation coefficients, F-score and a cross entropy based algorithm. It is found that all 4 filtering methods perform equally well. Moreover, we show that the worst performing algorithms are not detrimental to the overall performance

Using real-valued meta classifiers to integrate binding site predictions

Currently the best algorithms for transcription factor binding site prediction are severely limited in accuracy. There is good reason to believe that predictions from these different classes of algorithms could he used in conjunction to improve the quality of predictions. In this paper, we apply single layer networks, rules sets and support vector machines on predictions from 12 key real valued algorithms. Furthermore, we use a window of consecutive results in the input vector in order to contextualise the neighbouring results. We improve the classification result with the aid of under- and over- sampling techniques. We find that support vector machines outperform each of the original individual algorithms and the other classifiers employed in this work. In particular they have a better tradeoff between recall and precision.

Towards computational neural systems through developmental evolution

The capability of creating artificial neural networks with biologically-plausible characteristics, is becoming ever more attainable through the greater understanding of biological neural systems and the constant increases in available desktop computing power. Designing and implementing such neural systems still however remains a complex and challenging problem. This chapter introduces a design methodology, inspired by embryonic neural development, which is capable of creating 3 dimensional morphologies of both individual neurons and networks. Examples of such morphologies are given, which are created by evolving the parameters of the developmental model using an evolutionary algorithm.

Developmental Evolution of an Edge Detecting Retina

The task addressed in this paper is the evolution of an artificial retina with an on-centre/off-surround response which performs edge detection. Evolutionary optimisation is performed on the parameters of a developmental model. The model is capable of creating three dimensional, multilayer neural networks by modelling the outgrowth of neuron-to-neuron connectivity. A genetic algorithm is used to optimise the developmental parameters, measured against a target retina structure. The first stage of evolution adapts the parameters of outgrowth rules and the developmental environment. We show that this type of development can be sensitive to noisy conditions (perturbations in neuron positions). This limitation can be overcome by incorporating overgrowth and pruning. Staged evolution of these processes is shown to result in robust development.

Designing Development Rules for Artificial Evolution

Using artificial evolution to successfully create neural networks requires appropriate developmental algorithms. The aim is to determine the least complex set of rules that allow a range of networks to evolve. This paper presents a set of generic growth rules that abstractly model the biological processes associated with the development of neuron-to-neuron connections. Substantially different 3D artificial neural structures can be grown by changing parameter values associated with the rules. A genetic algorithm has been successfully employed in determining parameter values that lead to specific neural structures.

Using Real-Valued Meta Classifiers to Integrate and Contextualize Binding Site Predictions

Currently the best algorithms for transcription factor binding site predictions are severely limited in accuracy. However, a non-linear combination of these algorithms could improve the quality of predictions. A support-vector machine was applied to combine the predictions of 12 key real valued algorithms. The data was divided into a training set and a test set, of which two were constructed: filtered and unfiltered. In addition, a different window of consecutive results was used in the input vector in order to contextualize the neighbouring results. Finally, classification results were improved with the aid of under and over sampling techniques. Our major finding is that we can reduce the False-Positive rate significantly. We also found that the bigger the window, the higher the F-score, but the more likely it is to make a false positive prediction, with the best trade-off being a window size of about 7.

Evolving computational neural systems using synthetic developmental mechanisms

Original chapter can be found at: http://www.sciencedirect.com/science/book/9780124287655 Copyright Elsevier Ltd. [Full text of this item is not available in the UHRA]

A finite state automaton model for multi-neuron simulations

Abstract ‘Classical’ compartmental modelling techniques involve numerical solution of large sets of ordinary differential equations (ODEs). The computational cost of neuronal network evaluation using these techniques forms a major bottleneck in the exploration of the effects of dendrite morphology on network functionality. To reduce computational load, we have developed a finite-state automaton model of membrane activity, which will potentially permit the evaluation and comparison of large numbers of simulated 3-dimensional networks. The automaton mimics the behaviour of 2-equation ODE models for wave propagation in excitable media, and was found to be capable of modelling the most important characteristics of neural membranes.

Design of spatially extended neural networks for specific applications

The processes and mechanisms of biological neural development provide many powerful insights for the creation of artificial neural systems. Biological neural systems are, in general, much more effective in carrying out tasks such as face recognition and motion detection than artificial neural networks. An important difference between biological and (most) artificial neurons is that biological neurons have extensive treeshaped neurites (axons and dendrites) that are themselves capable of active signal transduction and integration. We present a model, inspired by the processes of neural development, which leads to the growth and formation of neuron-to-neuron connections. The neural architectures created have treeshaped neurites and contain spatial information on branch and synapse positions. Furthermore, we have prototyped a simple but efficient way of simulating signal transduction along neurites using a finite state automaton (FSA). We expect that the combination of our neuronal development method with the FSA that mimics signal transfer provide an efficient and effective tool for exploring the relationship between neural form and network function.