Analysis of Mitochondrial and Chloroplast Targeting Signals by Neural Network Systems

Abstract

Targeting sequences of nuclear-encoded mitochondrial and chloroplast proteins were investigated for characteristic sequence features. To evaluate the applicability of different neural network systems for feature extraction we employed a supervised and an unsupervised network training algorithm for targeting sequence analysis. Sets of sequences with less than 30% pairwise sequence identity were used in the experiments. Several physicochemical amino acid properties were used for sequence encoding. It turned out that the properties “refractivity”, “volume”, “polarity”, and “hydrophobicity” are suited for separation of chloroplast and mitochondrial targeting sequences. Prediction accuracy of the neural networks for separation of chloroplast and mitochondrial sequences yielded correlation coefficients around 0.7 in both training and test set. The sequences were encoded only by their mean property values, which were obtained from averaging over the first 20 residues of the precursors. To locate possible targeting signals in the precursor sequences Kohonen networks were used. These systems were able to identify several characteristic patterns of chloroplast and mitochondrial targeting sequences. The predominant role of the distribution of arginines in mitochondrial sequences is substantiated by our findings. Chloroplast sequences seem to be characterized by stretches containing high contents of alanine, serine, and threonine. Putative locations of the targeting signals were found using the Kohonen networks for prediction of the features extracted. Analysis of the FNR precursor from Cyanophora paradoxa served as an example.