Reconhecimento de padrões e redes neurais artificiais em predição de estruturas secundárias de proteínas

Abstract

This thesis proposes to use pattern recognition and MultiLayer Perceptron (MLP) artificial neural networks in protein secondary structure prediction. We get the latest improvements of previous works, namely: variation in the size of windows; use of evolutionary information; use of decision jury and creation of neural networks chain and propose: bringing together all these improvements in a single tool (cMLP), to test with a set of proteins already tested by previous works and to perform the comparison of the tool built with other classifiers already available on the web, all arising from papers, and thus perform an optimization process of the neural network parameters. The conventional neural network with random initializations was optimized, becoming an optimized MLP neural network (oMLP). The optimization process occurred in the initialization parameters of the neural network project based in the multi-class linear discriminant analysis and the weighted Fisher Criterion subspaces extraction, even in the selection input data. We hope that the oMLP network will be effective in the process of dimensionality reduction of the search space in extensive groups of protein data will enhance the most of the performance measures by including the accuracy of the secondary structure prediction and convergence properties.