Pattern recognition of steroids using fragment molecular connectivity.

The use of pattern recognition methods to classify a set of steroids into five therapeutic categories was investigated. First-order fragment molecular connectivity values were determined for 10 positions on each molecule using a template-based method of position assignment. Learning set and test set classifications were performed. Although the numbers of compounds misclassified were comparable for all of the methods, the identities of the misclassified compounds varied depending on whether the classification method assumed a local or a global view of the data. The best classification results were comparable to those obtained by linear and quadratic discriminant analyses. For this set of compounds, it was concluded that pattern recognition methods offer no advantages over traditional discriminant analysis methods if classification alone is considered, especially since most discriminant analysis procedures utilize stepwise variable selection, which is not as common in pattern recognition analyses.