A Probabilistic Inference System for The Prediction of Subcellular Localization Sites of Proteins: Application to E. coli Data

We have proposed that the prediction of protein subcellular localization sites can provide a good clue for the characterization of open reading frames of unknown function [1, 2]. Our program, PSORT, has been used by a number of researchers through the Internet [3]. PSORT was originally written in the style of a 'if-then' rule-based system. Although this style has the merit of a great versatility in coding inference pathways, re-optimization of numeric parameters for either given training data or expanded rules needs an expert's manual work. Clearly, this character is not well suited for the rapidly-progressing state of genome analyses. Thus, we have been studying other mathematical models for inference that allow at least semi-automatic optimization of numeric parameters. Last year, we introduced a simple model, the waterow model, that automatically nds required threshold parameters [4]. This model showed su ciently high discrimination power for model data. Here, we describe an improved model and report its predictability when applied to more realistic data. We rst collected E. coli amino acid sequences of known subcellular localization sites from the PROSITE database (Rel. 31). Excluding hypothetical information, 336 sequences were collected in total. They were classi ed into the following 8 groups: lipoproteins at the inner membrane (imL), lipoproteins at the outer membrane (omL), inner membrane proteins with a cleavable signal sequence (imS), typical outer membrane proteins (om), periplasmic proteins (pp), inner membrane proteins with a signalanchor 2Cf0f8,B@!’Bg:eBg3X:YK&@8BN9)3X%;%s%?!<!$") 565 ?aED;T;3ED5V 1-3