Copyright © 2003 BioChem Press Classification of Polar and Nonpolar Aquatic Pollutants Using Simple Descriptors. Differences between Polarity Prediction

Motivation. The problem of toxicity prediction is mainly related to the necessity of processing many data that most of the time come from different sources and have different biological meaning. Often, the real mechanism of action of a toxicant is unclear or difficult to reproduce; in addition, a chemical compound exercises its toxic action through many steps that depend both on its structure and on the specific environment where it is acting. In this perspective, the classification of compounds can be of great help because decreases the number of the alternatives to those specific of that class, allowing a more focused analysis. The classification of narcotic pollutants into polar and nonpolar sets is certainly an important aspect of this type of problems. Method. Object classification requires two principal components: the selection of the molecular descriptors and the choice of the classification algorithm. The calculation of the molecular descriptors is performed using our own approach that is based on empirical equations. We calculated three descriptors (Helc, HQ+, Elcdif) that are used in pairs (Helc and Elcdif, or HQ+ and Elcdif). Using two classification algorithms, a classical neural network and a tree neural network, we analyze two compound sets; the first contains 190 narcotic pollutants (114 nonpolar and 76 polar), the second contains 30 pollutants (20 nonpolar, 5 polar, 5 acetylcholinesterase inhibitors). In a broad sense, the first set is used as training set and the second as test set