A general algorithm is presented for identifying sets of positions in multiple sequence alignments that best characterize an a priori partitioning such as those determined by inhibition studies or other experimental techniques. The algorithm explores combinations of polymorphic columns in the alignment and evaluates how well these sites reflect the original input partition. Partitions across the polymorphic columns are derived using a tree building procedure with conventional amino acid substitution matrices. Elucidation of those amino acids which govern the biochemical behaviour of a protein with a given substrate or inhibitor can provide insights towards an understanding of the tertiary conformation of the protein. Since it is likely that such positions will be spatially clustered in the protein fold, these positions may give rise to useful distance constraints for substantiating model protein structures. The method is exemplified using data for a set of human mu class glutathione S-transferases. A novel aspect for predicting the behaviour of new polymorphic sequences is also discussed.