Which rules for assembling short-term test batteries to predict carcinogenicity?

The main theme of this paper is to describe the basic requirements for assembling reliable batteries of short-term tests for carcinogenicity prediction. For this purpose, a subset of the data base generated by the International Program for Evaluation of Short-Term Tests for Carcinogens (IPESTTC) has been studied by different data-analysis methods. Much attention has been focused on the methodological dimensions of the problem, at the level of selection of both data and statistical techniques. Twenty-one of the most widely used short-term assays were considered. An exploratory study of the data base was first performed by factor analysis, showing similarities and dissimilarities between test performances and confirming our previous results obtained by cluster analysis. In this way the assays were divided into three groups on the basis of their responses to the chemicals. The Salmonella assay was in the central group, characterized by equilibrated performances in respect to sensitivity and specificity for carcinogens. Tests complementary to the Salmonella assay for sensitivity and specificity, respectively, were identified as well. A preliminary comparison of the IPESTTC results with the Gene-Tox data base is also presented. The test performances in respect to carcinogenicity prediction were then evaluated by discriminant analysis. When the subset of data was considered as a whole, the procedure resulted in a linear discriminant function able to correctly identify 84.2% of carcinogens and 83.3% of noncarcinogens. The correctly identified carcinogens summed to about 90% when adequate batteries of tests were used. This analysis yielded a number of observations. (1) Together with the selectivity indices (such as sensitivity and specificity), the operational complementarity between test performances must be ascertained and taken into account. (2) The batteries most effective at predicting carcinogenic activity were composed of three tests, one for each group. This finding converged with the fact that the three classes of assays were clearly differentiated for sensitivity and specificity, and in this sense were complementary to each other. (3) The performances of the batteries were not improved by adding more tests, but in several cases the opposite effect occurred. (4) Estimation of the probability of the chemicals of being carcinogenic, starting from qualitative genotoxicity data, is possible.