Table 3: Examples of in silico models for mutagenicity.

Model	Example	Design
Knowledge- or rule-based Structure-activity relationship (SAR)	Toxtree (public model)	Toxtree includes modules for mutagenicity, carcinogenicity, and the in vivo micronucleus assay Serafimova [55] The model includes a decision tree for assessment of mutagenicity and carcinogenicity potential by discriminant analysis and structural rules Benigni [62] The model also includes a decision tree for the in vivo micronucleus assay. The accuracy of prediction is 70% for carcinogenicity, 78% for mutagenicity and 59% for the in vivo micronucleus assay.
	Derek Nexus (Commercial model)	This (rule-based) model contains structural alerts for mutagenicity, chromosome damage and carcinogenicity. The hazard assessment is justified with literature references. Advantages are the transparency in predictions, the rule development is peer-reviewed by a user group, and new rules can be added easily. The absence of a predicted hazard simply means that no relevant alerts were identified; it does not necessarily mean the absence of hazard Serafimova [55].
Quantitative or statically based Structure-activity relationship (QSAR)	CASE Ultra (Commercial model)	This model applies a statistical approach that automatically identifies molecular substructures that have a high probability of being relevant to the observed endpoint. Genotoxicity models include Ames mutagenicity, direct mutagenicity, base-pair mutagenicity, frameshift mutagenicity, chromosomal aberrations, mouse micronucleus assay, mouse sister chromosomal exchange. Carcinogenicity models include rat, mouse, female, male carcinogenicity, TD50 rat, mouse carcinogenicity Serafimova [55].
	VEGA (Public model)	A statistical model for mutagenicity that was originally developed in the frame of the EU CAESAR project. The authors reported correct classification rates of 92.3% and 83.2% for the training and test sets, respectively. The model was combined with Toxtree, showing that the number of false negatives could be reduced, but the number of the false positives increased. The authors concluded that by using the so-called “cascade model”, a classification accuracy close to the reliability of the Ames test data could be achieved Serafimova [55]. In the same project, two complementary approaches (regression and classification) were applied to develop models for carcinogenicity and an accuracy of classification of 91-96% for the training set and 68-74% for the test set was reported Serafimova [55].