Intelligent programs for structure elucidation should be able to predict structural features, or even complete structures, for an unknown compound. This can be accomplished by substructure analysis of the topological matrices corresponding to the most similar reference spectra detected by a library search. A comparison of the results from different spectroscopic techniques ('H NMR, uC NMR, IR, MS) reflects the content of the individual data bases, and also the sensitivity of the specific physical experiment to structural changes. The results can be combined to yield multidimensional information on the probability of any given substructure, even without crosslinked libraries from varying sources. In connection with the structure generator and the uC NMR spectrum estimation in ACCESS, a prediction of novel structures is routinely possible, as exemplified by the analysis of the spectra of 3'-methylacetophenone and genipin. The great advantages of this approach lie in its high flexibility in omitting or adding information and the absence of hand-encoded rules, because knowledge of the system simply stems from its experience, i.e. the structure-oriented spectral data bases.

No abstract