Scientist at Universities and in Industry are producing huge amounts of scientific information. Especially in Chemistry a considerable amount is presented in a graphical manner, e.g., structural representations, reaction schemes and spectra.In many topics of modern chemical research the representation of a chemical structure, e.g., as a simple ball&stick model, is not sufficient to describe all the properties that are related to the structure. Thus, the chemist should have an access to graphics oriented tools for the retrieval or production of data and to visualize the new information in graphical manner.
TeleSpec is one of the premier projects that aid this task. TeleSpec deals with new techniques of molecular representations as well as new techniques of data processing and makes them available via Internet. The aim of the TeleSpec project was to provide an information pool for infrared spectroscopy, and to present sophisticated techniques for the analysis of infrared spectra.
Substance identification by IR spectroscopic methods is usually performed by comparing an experimental spectrum with a reference spectrum from a spectrum library. This identification technique assumes that a reference spectrum for the query spectrum is available. The high discrepancy between the amount of about 17.000.000 known chemical compounds and about 100.000 spectra stored in the largest infrared spectra database often prohibits this easy way of substance identification. A method is presented based on a combination of a neural network with a novel structure coding scheme that allows the rapid simulation of infrared spectra and thus supplies access to reference spectra for arbitrary query molecules independent of their size.
Within the scope of the TeleSpec-project this spectrum simulation method is made available via the Internet. The user can perform interactive spectra simulation experiments and interpret the results by online analysis of the neural network.
The relationship between structures is quite complex and cannot be described by a simple equations. Neural networks are very useful tools for modelling nonlinear correlations. The neural networks learn from experimental data about the correlation of structure and spectrum by analyzing a set of examples. By presenting the query data to a trained neural network, the learned correlation can be used to predict an IR spectrum for an unknownm compound.