Correlations between validation and definitive study results for genotoxic compounds.

Abstract

The use of batteries of predictive tests, perhaps in lieu of costly definitive testing, requires thoughtful identification of individual assays and definition of test batteries and tiers. Generally predictive assays are appraised in a validation study which uses a number of compounds known to be positive (or negative) in the definitive assay. A predictive test meeting specified criteria of specificity (proportion of true positives detected) and sensitivity (proportion of true negatives detected) is then eligible for use in screening compounds of unknown potency. We formally derive the definitions of prevalence (proportion of true positives in a sample of N compounds), sensitivity, and specificity, and show that these arise from different multinomial distributions for validation and screening studies. The formal statistical correlations between the predictive tier and the definitive tier are derived for both types of studies. Correlations show that a predictive assay which is eliminated in a validation study because of low sensitivity or specificity may perform satisfactorily when used for screening. Conversely, a predictive assay which performs well in a validation study may have a poor correlation with definitive assay results when used for screening.