compared with the Ashby’s SAs. Moreover, modulating factors can
inhibit or cancel the toxic effects of SAs [83]. In 2012, Wang et al.
developed a model by building and pruning a molecular fragments
tree to select high-quality SAs for carcinogenicity [84]. Finally,
77 SAs and 4 modulating factors produced higher predictive ability
than Benigni’s model. As an effective measure in the evaluation of
genotoxicity and carcinogenicity, the identification of SAs attracts
much attention in the screening of drug candidates and toxicity
testing [85].3.2 Local Models Most of the genotoxic carcinogens generally have the unifying
feature that they are electrophiles or can be activated to electro-
philic reactive intermediates [74]. Multiple QSTR models have
been developed for numerous congeneric series of genotoxic carci-
nogens, such as aromatic amines, nitroaromatic compounds,N-
nitroso compounds, quinolines, triazenes, polycyclic aromatic
hydrocarbons, and halogenated aliphatics [74]. However, for non-
genotoxic carcinogens, the QSTR predictions are still scarce due to
their complex mechanisms of carcinogenesis [86].
3.3 Global Models Global models are useful for predicting noncongeneric classes of
chemicals that have diverse chemical scaffolds and complex
mechanisms of carcinogenicity [76]. In general, the global models
perform inferior to the local models because the global models
consider several mechanisms of action at the same time
[74]. According to the result of predictive toxicity challenge in
2000–2001, only 5 out of 111 models for classification performed
better than random guessing [87]. In 2003, Contrera et al. con-
structed a MDL QSAR model using molecular structural similarity
and E-state indices and had excellent coverage (93%) and good
sensitivity (72%) and specificity (72%) for rodent carcinogenicity
[51]. In 2004, Sun et al. developed a PLS-DA (partial least squares
discriminant analysis) model for predicting carcinogenicity,
showing R^2 ¼ 0.987 and Q^2 ¼ 0.944 for male mouse,
R^2 ¼0.985 andQ^2 ¼0.950 for female mouse,R^2 ¼0.989 and
Q^2 ¼0.962 for male rat, andR^2 ¼0.990 andQ^2 ¼0.965 for
female rat [88]. Moreover, Tanabe et al. used ensemble learning
technique to divide the data set into 20 subsets based on the
contained substructures and built SVM models for each subset
with an overall accuracy of approximately 80% [89]. In 2013,
Singh et al. established the classification model using probabilistic
neural network and the regression model using generalized regres-
sion neural network based on 834 structurally diverse chemicals
from CPDB (Carcinogenic Potency Database) [90]. Both models
exhibited excellent prediction ability, which are valuable for safety
evaluations of chemicals. In 2015, Li et al. constructed the binary
(carcinogen and non-carcinogen) and ternary (strong, weak carcin-
ogen and non-carcinogen) classification models using six types of
254 Jing Lu et al.