next up previous
Next: Measure of Accuracy Up: SPEP: a Signal Peptide Previous: SPEP: a Signal Peptide

Introduction

The detection of signal peptides is of great relevance in the field of automatic genomics annotation. Several methods have been developed to predict the presence of the signal peptides and their cleavage sites [*]. Starting from the von Heijne's pioneering paper [12] to the most advanced machine learning approaches, which include Support Vector Machines (SVM) [11], Hidden Markov Models (HMM) [9] and hybrid methods (HMM and Neural Networks) [10]. We implemented two types of neural networks, one for the detection of the cleavage sites (CleavageNet) and the other for the N-terminus discrimination (SignalNet). Although our architectures are similar to those previously described by other authors [10], our training was carried out using a new data set of experimentally determined signal peptides [7]. Moreover we implement a different way of combining the two networks, reaching the same level of accuracy achieved before [10], without using hidden Markov models. We train three different predictors: for gram negative, for gram positive and for Eukaryotes. In all cases we adopted a cross validation procedure (taking care of eliminating sequence with detectable homology among training and testing sets). The results for the best performing predictors are reported below, together with the application of the prediction to Escherichia coli annotated proteome.
next up previous
Next: Measure of Accuracy Up: SPEP: a Signal Peptide Previous: SPEP: a Signal Peptide
2003-06-12