A computer-assisted pattern-recognition system (ADAPT) designed to elucidate structure-activity relationships was applied to a set of retinoids, potentially useful inhibitors of carcinogenesis. A data set of 67 retinoids was used as input to the ADAPT system; their structures were entered, and their 3-dimensional configurations were optimized by a molecular modelling algorithm. Forty of these retinoids were defined as the "more active" class based upon their ability to reverse keratinization in vitamin A-deficient hamster tracheal organ cultures at 10(-10) M or less. The remaining 27 retinoids were defined as the "less active" class due to their lack of ability to elicit this effect at 10(-8) M or more. Thirteen descriptors were generated by ADAPT for each of these retinoids based upon their structures, including: number of ring atoms; double bonds; del Ré sigma charges; and principal moments. Pattern recognition analysis techniques were applied to this data set to determine if information contained in these descriptors could generate a discriminant function equation which could separate more active from less active retinoids, successfully. Computer recognition of more active from less active retinoids was demonstrated by a number of pattern recognition techniques, and the discriminant function could predict correctly the relative activity of retinoids of "unknown" activity in 87% of trials. These results indicate the existence of distinct structure-activity relationships in this set of biologically important molecules.