Assessment of the sensitivity of the computational programs DEREK, TOPKAT, and MCASE in the prediction of the genotoxicity of pharmaceutical molecules

Computational models are currently being used by regulatory agencies and within the pharmaceutical industry to predict the mutagenic potential of new chemical entities. These models rely heavily, although not exclusively, on bacterial mutagenicity data of nonpharmaceutical‐type molecules as the primary knowledge base. To what extent, if any, this has limited the ability of these programs to predict genotoxicity of pharmaceuticals is not clear. In order to address this question, a panel of 394 marketed pharmaceuticals with Ames Salmonella reversion assay and other genetic toxicology findings was extracted from the 2000–2002 Physicians' Desk Reference and evaluated using MCASE, TOPKAT, and DEREK, the three most commonly used computational databases. These evaluations indicate a generally poor sensitivity of all systems for predicting Ames positivity (43.4–51.9% sensitivity) and even poorer sensitivity in prediction of other genotoxicities (e.g., in vitro cytogenetics positive; 21.3–31.9%). As might be expected, all three programs were more highly predictive for molecules containing carcinogenicity structural alerts (i.e., the so‐called Ashby alerts; 61% ± 14% sensitivity) than for those without such alerts (12% ± 6% sensitivity). Taking all genotoxicity assay findings into consideration, there were 84 instances in which positive genotoxicity results could not be explained in terms of structural alerts, suggesting the possibility of alternative mechanisms of genotoxicity not relating to covalent drug‐DNA interaction. These observations suggest that the current computational systems when applied in a traditional global sense do not provide sufficient predictivity of bacterial mutagenicity (and are even less accurate at predicting genotoxicity in tests other than the Salmonella reversion assay) to be of significant value in routine drug safety applications. This relative inability of all three programs to predict the genotoxicity of drugs not carrying obvious DNA‐reactive moieties is discussed with respect to the nature of the drugs whose positive responses were not predicted and to expectations of improving the predictivity of these programs. Limitations are primarily a consequence of incomplete understanding of the fundamental genotoxic mechanisms of nonstructurally alerting drugs rather than inherent deficiencies in the computational programs. Irrespective of their predictive power, however, these programs are valuable repositories of structure‐activity rela