The human brain generates maps of the environment from sensory information. This capability of the human brain is modelled by self-organizing neural networks such as the one developed by Kohonen. Kohonen networks can be used for the mapping of molecular surface properties. It is shown that maps of the molecular electrostatic potential provide valuable information for understanding biological activity and searching for new lead structures. Kohonen networks can also be used for the mapping of datasets of molecules. Autocorrelation vectors derived from the topology of molecules or from molecular surface properties provide an encoding of molecular structures that can be used as input to Kohonen networks and, thus, allow a clustering of molecules that reflects biological activity. Such a mapping can be used for the assessment of the similarity and diversity of chemical libraries.
The algorithms presented here, both those for the calculation of physico-chemical effects such as the molecular electrostatic potential and that for the Kohonen network, work quite rapidly. In addition, by their very nature, neural networks are of a parallel manner allowing their implementation on parallel machines. This all taken together makes it possible to study large molecules and very large datasets.