Oral Presentations Self organizing map and counterpropagation neural networks in structure-property modelling: Examples from environmental science and drug design

Self organizing map technique is often used to analyse the data in multi-dimensional space. Basis of this technique is a non-linear projection from multi-dimensional space onto two-dimensional grid of neurons (map). The projection, which is topology preserving but not metric preserving, is achieved via non-linear algorithm known as training. The fundamental property of the trained network is that the similar objects are located close to each other. Counterpropagation neural network is a generalization of self organizing map. Additionally, it takes into account the property (output) values. In the presentation the architecture and the learning strategy of both artificial neural network techniques are discussed (1,2). Self organizing map and counterpropagation neural network are suitable techniques in molecular structure-property relationship studies (3). They can be applied to build the predictive models, for classification, clustering, determination of outliers, or selection of most relevant descriptors (4). In the presentation we discuss different strategies of application of both methods in structure-property relationship studies. We present several examples where different data sets and different properties, which are interesting in environmental science (aquatic toxicity, mutagenic potency, log P) and in drug design (partitioning coefficients) have been studied. In addition, we discuss how to use the both techniques in analysis of 2-D electrophoreses (proteome) maps.