A statistical-heuristic method for selecting drugs for animal screening is developed with molecular structure features as predictors of biological activity. The method is intended to work on large amounts of data over varied structures. A trial of this method on a small data set allows some comparison with more sophisticated pattern recognition methods. Problems connected with interdependence among structure predictors are critical in this method and schemes to eliminate redundancy are reviewed. Alternate sets of structure predictors are considered. The discussion here outlines directions to be taken in the near future.

This paper reports the application of pattern recognition and substructural analysis to the problem of predicting the antineoplastic activity of 24 test compounds in an experimental mouse brain tumor system based on 138 structurally diverse compounds tested in this tumor system. The molecules were represented by three types of substructural fragments, the augmented atom, the heteropath, and the ring fragments. Of the two pattern recognition methods used to predict the activity of the test compounds the nearest neighbor method predicted 83% correctly while the learning machine method predicted 92% correctly. The test structures and the important substructural fragments used in this study are given and the implications of these results are discussed.

No abstract

No abstract

No abstract