Using a computer-aided approach, the tests for Salmonella mutagenicity and transformation in established cell lines were compared for the qualitative bases of their carcinogenicity predictions. For this purpose, a database of 145 chemicals was prepared in which rodent carcinogenicity data and results of the Ames' and transformation tests were available. Using a software program for connectivity analysis (previously developed and validated by us), we assayed the molecular structures of these chemicals for the presence of fragments relatable to their positive (i.e., biophores) or negative (i.e., biophobes) response to the tests in question. These fragments were then studied for their association with genotoxic and nongenotoxic carcinogenicity. The philosophy adopted was that the type and number of molecular fragments chosen by the software to describe the chemicals correctly predicted by the tests could be related to the type of carcinogenic effects to which the tests themselves were sensitive. The classifications made by the software were interpreted by human expertise and the biophores found were compared with the acknowledged structural alerts to DNA reactivity as formalized by Ashby and co-workers [(1991): Mutat Res 257:229-306; (1993): Mutat Res 286: 3-74]. The results show that, in quantitative terms, the overall ability to predict carcinogenicity is about the same for both the Salmonella and transformation tests. However, in qualitative terms the transformation test appears to be sensitive to effects that are more heterogeneous than those inducing mutation, some of which are presumably related to nongenotoxic carcinogenic activities. This study illustrates a possible, innovative model of analysis of chemical structures that, using an automated approach along with the biologist's judgment, could contribute to the detection of complementarities among short-term test endpoints.