Atomistic topological indices applied to benzodiazepines using various regression methods

The use of atomic positionally independent molecular representations of a set of benzodiazepines is shown to yield a comparable structure-activity model than the traditional physicochemical representations. The best model was produced using an artificial neural network; the results are compared with models produced by MLR, PCR and PLS. The atomistic representation compares favourably with a functional group representation and the idea of implicit information in such representations is discussed. It appears that the complex interrelationships between elements of a simple molecular representation may be capable of encoding subtle molecular structural elements.