The knowledge-based expert computer system DEREK (Deductive Estimation of Risk from Existing Knowledge) has been assessed for its potential as a screen for predicting genotoxicity and carcinogenicity of some chemicals found in foods. This was achieved by establishing databases, summarizing published activities of several chemical classes, including aflatoxins, flavonols, hydroxylated anthraquinones and polycyclic aromatic hydrocarbons (PAHs), to construct draft rules to identify structurally alerting toxicophores. This information was used, together with other data (e.g. physicochemical properties, steric effects and molecular modelling), to produce a set of simplified, non-specific rules where possible to cover a wide spectrum of structures within each chemical class. Examples of toxicophores identified include: a bisfuranoid substructure centred around the 2,3-vinyl ether bond for aflatoxins and a 6-bond substructure, based on the 'bay' region of phenanthrene, for PAHs. The effects of various substituents and the presence of extra rings were considered. Operation of the rules is illustrated by reference to these chemical groups, and to the flavonols for which there are four on-screen messages, depending on hydroxylation at positions 3 and 5 on the fused ring system and on the phenyl appendage, and whether further hydroxylation can occur, for example by mixed function oxidase activity and/or chemical or metabolic release of masked hydroxyls. Rules were tested by processing a wide range of structurally related chemicals in each class, with known and unknown activities. In some cases, key candidate structures for synthesis and toxicity testing were identified. The results from such studies should facilitate rule improvement. Whereas predictivity of a set of rules depends on the availability of reliable, experimental toxicity data for as many potential molecules in a group as possible, utility for identifying toxic untested, novel structures may be enhanced by using physicochemical information. Examples of this approach, as well as the advantages and current limitations of DEREK, are presented and discussed. It is concluded that DEREK has potential as a screen for genotoxicity and the approach adopted here for rule development may be useful for predicting other toxicity endpoints.