Conformationally constrained environment-dependent amino acid residue substitution tables have been constructed from a database comprising 33 homologous families of protein sequences aligned on the basis of their three-dimensional structures. Residues are allotted to one of 216 (or 54) classes of combinations of structural features. These include nine main-chain conformation classes, three classes of side-chain accessibility and eight (or two) classes of side-chain involvement in three types of hydrogen bond. Seven different main-chain conformational classes outside of regions of regular structure were identified in an analysis of the distributions of phi-psi torsion angles in 84 high-resolution crystallographic structures. Residue substitutions at equivalent positions in the structural alignments are included where the main-chain conformational class is conserved. Frequency data in the form of 216 (or 54) environment specific (20 x 20 residue type) matrices are then converted to probabilities. Two smoothing regimes incorporating entropy-driven weights were applied to the set of 54 tables. Predicted residue substitutions have been generated for individual residue positions in beta-hairpins and the hypervariable regions of the immunoglobulins. These have been compared with the observed sequence variation at the same positions using rank correlation methods. Measurements of chi 2 distances demonstrate the considerable improvement in predictive power at key residue positions identified from interactive graphics studies when compared to the Dayhoff MDM250 mutation matrix. An illustrative example is given of an application of the method in the ranking of loop fragments in model building studies of structurally variable regions in two subtilisins. A combined template scoring procedure is found to be 26-fold more discriminatory than the Dayhoff matrix. The success rate is approximately 85%.