Improving

signal peptide prediction accuracy by simulated neural network I.Ladunga1, F.Czakd2, I.Csabai2 and T.Geszti2 The accuracy of distinguishing amino-terminal signal peptides from cytosolic proteins has been improved to 95 % by combining a neural network classifier with von Heijne's statistical prediction, the latter is itself 85-90 % reliable. The network processed not the cleavage site, but amino-terminal 20-residue segments by the 'tiling ' algorithm. Concordant positive predictions of both methods led to the safe identification of 496 novel signal peptides from the Protein Identification Resources. Signal peptides are short—mostly amino-terminal—segments targeting proteins across membranes (Gierasch, 1989), then split off the preprotein as the target has been reached. Rules (or rather trends) specifying cleavage site patterns have been deduced (Perlman and Halvorson, 1983; von Heijne, 1983) from a limited number of signal sequences determined by peptide chemical methods, by ultracentrifugation of the membrane or by immunoelectron microscopy (Randall etal, 1987). Statistical analysis of preferred residues in given positions yielded predictions about whether a given sequence, as scanned by a sliding window spanning over 12 or 15 residues, would contain a signal peptide cleavage site (von Heijne, 1986; Fol