Validation of the (Q)SAR combination approach for mutagenicity prediction of flavor chemicals

Abstract Most exposure levels of flavor in food are considered to be extremely low. If at all, genotoxic properties should be taken into account in safety evaluations. We have recently established a (quantitative) structure–activity relationship, (Q)SAR, combination system, which is composed of three individual models of mutagenicity prediction for industrial chemicals. A decision on mutagenicity is defined as the combination of predictive results from the three models. To validate the utility of our (Q)SAR system for flavor evaluation, we assessed 367 flavor chemicals that had been evaluated mainly by JECFA and for which Ames test results were available. When two or more models gave a positive evaluation, the sensitivity was low (19.4%). In contrast, when one or more models gave a positive evaluation, the sensitivity increased to 47.2%. The contribution of this increased sensitivity was mainly due to the result of the prediction by Derek for Windows, which is a knowledge-based model. Structural analysis of false negatives indicated some common sub-structures. The approach of improving sub-structural alerts could effectively contribute to increasing the predictability of the mutagenicity of flavors, because many flavors possess categorically similar functional sub-structures or are composed of a series of derivatives.