General Information Network Training Decoding of an RDF Code

The Training of a CPG network

A database of infrared spectra (SpecInfo database from Chemical Concepts containing about 13,500 infrared spectra) is searched for those 50 spectra that are most similar to the query infrared spectrum from which the structure should be derived. Similarity of infrared spectra is determined by the correlation coefficient between a database spectrum and the query spectrum. The database contains both, infrared spectra and the connection tables of the corresponding molecules. In a second step, 3D molecular models are generated from the connection tables by the structure generator CORINA for the structures selected on the basis of similarity of their infrared spectra. After the calculation of physicochemical atomic properties the 3D molecular models are encoded as discrete values of the atom radial distribution function (RDF) using the program ARC. Then the counterpropagation network is trained with the selected structures encoded by the RDF code and their infrared spectra. The query infrared spectrum is entered into the CPG network. The structure code found in the winning neuron represents the output that has to be decoded to obtain the 3D structure.