SAR model development should be a continuous process involving formulation then experimental testing of the model, incorporation of test results into the database, then re®nement of the model using the expanded database. The larger database affords greater con®dence in its ability to predict the biological response. This iterative procedure was employed with a recently developed structure-activity relationship (SAR) model of human skin irritation. Based on a ``leave-one-out'' cross validation, the mean sensitivity of the initial model was 0.89, the mean speci®city was 0.74. A clinical validation study was conducted to assess the ability of the model to predict human skin irritation by esters commonly used as fragrance ingredients. Esters that were found to cause irritation in rabbits, and that were within the predictive space of the SAR model, were selected for human testing using the patch test procedure. Of the 34 rabbit irritants selected, 16 were predicted by the model to be positive and 18 were predicted to be negative. Patch testing yielded two positive esters, allyl heptanoate and allyl cyclohexanepropionate. These test results were incorporated into the database to re®ne the SAR model. Best subsets regression and linear discriminant analysis were used to generate 10 submodels consisting of 10 irritants and 50 non-irritants randomly selected from the new database. Physicochemical parameters associated with irritant esters, when compared with non-irritant esters, differed somewhat from those identi®ed in the original model. Irritant esters had lower solubility parameter and water solubility, higher Hansen dispersion and Hansen hydrogen bonding, and lower sum of partial positive charges, when compared with non-irritant esters. The sensitivity of the new model is 0.69 and speci®city is 0.67. The results of this study indicate that SAR models based on limited data may not accurately predict the activity of unknown chemicals even though the computationally-derived sensitivity and speci®city of the models are high. This ®nding emphasizes the need for experimental validation of models and their re®nement as new data become available.