Approaches to the development of structure-activity relationship (SAR) models for dermal irritation

The goal of this research was to develop structure-activity relationship (SAR) models that could predict irritation based on chemical structure and physicochemical properties to help understand the underlying mechanisms and prevent disease. The approach involved database development, model building, testing, and refinement. Model testing and refinement ensure that SAR models are applicable to untested chemicals.;A diverse human database of 55 irritants and 82 non-irritants was developed from a literature review. A diverse database of 104 rabbit irritants and 56 non-irritants was also created.;CASE/MultiCASE modeling identified several biophores associated with human irritants. However, the biophores were found only in about half of the irritants. CASE/MultiCASE modeling of subsets of the rabbit database identified one structural alert consistently associated with irritation. None of the alerts were in common with the human model.;Physicochemical models of the human database focused on esters, the largest subset of chemicals. Physicochemical modeling of the rabbit irritants failed to accurately predict human irritation. To increase the applicability of the human SAR model, multiple random sampling was used to create subsets of irritants and associated SAR models. From these, models, consensus models were created consisting of the properties that were statistically significant and consistently associated with irritation. An initial SAR model was developed that had good predictivity based on cross-validation (sensitivity = 0.89, specificity = 0.74). The model was tested in a clinical study using 34 esters not previously tested in humans. The model correctly predicted activity of 20 esters. The remainder were incorrectly predicted, indicating that a model based on limited data may not accurately predict new chemicals. The model was refined by incorporating the clinical results into the database. The refined model had a sensitivity of 0.69 and a specificity of 0.67 when used as a consensus model. The physicochemical properties associated with irritation indicated that partitioning into the lipid-rich epidermis is likely an important factor in skin irritation of these esters. In addition, an interaction such as receptor binding may also be involved. The model should prove useful in predicting the irritation potential of untested esters and eventually preventing response in the public.