Several methods have been described in the literature for the calculation of weights which represent the contribution of fragments to the overall activity or inactivity of molecules which contain them. These weighting schemes are based on fragment occurrence data in training sets of molecules for which the activity is known. This paper reports a comparison of several such schemes (fourteen fragment weights in all), using small datasets for which structural and activity data are available. The comparison reveals that the most effective weight seems to be one derived from research into document retrieval systems (where indexing terms are used to discriminate between relevant and non‐relevant documents).

This paper considers the use of fragment weighting schemes for substructural analysis studies. Experiments are reported using 2-D and 3-D substructural descriptors with both small and large datasets for which qualitative activity data are available. The results support previous work in suggesting the potential of probabilistic models of weighting first used in studies of text retrieval.to whom all correspondence should be addressed