Applications of High-Resolution Self-Organizing Maps to Retrosynthetic and QSAR Analysis

Kohonen's self-organizing map (SOM) is a neural network model of the unsupervised class and, by some aspects, is analogous to other clustering algorithms. The high-resolution maps are characterized by a far higher number of neurons than the number of learning patterns to deal with. This paper describes explorations of two sets of organic compounds with this technique. The first set includes 32 carbonyl derivatives encoded by molecular structural subunits and classified according to four disconnection pathways. The self-organizing map is able to separate the 32 derivatives into 4 clusters without a prior knowledge of class membership. The results obtained are compared with those of a supervised multilayer perceptron. The second experiment tests the self-organization of 64 hallucinogenic phenylalkylamines encoded with a list of substituent descriptors (hydrophobicity, volume, and C-13 NMR relative aromatic shift). The network organizes the hallucinogenic compounds on the map according to their structural similarities; they are also globally positioned in relation to their biological activities.