kNN, naive Bayes (NB), and C4.5 decision tree [60], were
employed for model building. In this study, one-against-one
(OAO) [61, 62] and binary tree (BT) [63, 64] strategies were
used to extend SVM for multiclass classification, respectively.
The results indicated that the OAO-SVM model not only achieved
the best performance for external validation set I and II but also
showed the good prediction ability for each class. Moreover, the
privileged fragments in Categories I (Danger/Poison) and II
(Warning) were extracted by substructure frequency analysis and
information gain methods, which can help us to quickly identify the
toxicity category of a new compound.2.3 Software and
Web Servers for Acute
Toxicity
Nowadays, multiple popular software and web servers can predict
acute toxicity following various administration routes in multiple
species. For example, ACD/Labs ToxSuite predicted LD 50 values
of the compounds by considering expert knowledge of various
physiological responses (e.g., inhibition of cholinesterase and ATP
synthesis) and structure-activity relationship [65]. Toxicity Estima-
tion Software Tool (TEST) employed the HC andkNN methods to
build individual models and took an average of the predicted values
as the final prediction [66]. Accelrys TOPKAT package developed
19 linear models for predicting LD 50 values of multiple classes of
compounds and therefore intuitively hinted the positive or negative
contributions of descriptors to the toxicity values of the specific
classes [67].
Gonella Diaza et al. employed five software programs, includ-
ing ACD/ToxSuite, TEST, TOPKAT, ADMET Predictor [68],
and TerraQSAR [69], to evaluate acute oral toxicity of 7417 com-
pounds [70]. Among these five models, TEST showed the best
performance, which yieldedR^2 of 0.74 for the training set and 0.60
for the test set. TOPKAT had inferior prediction ability (R^2 ¼0.61
for the training set and 0.34 for the test set).
ProTox is a web server for the prediction of oral LD 50 values in
rodents based on chemical similarity and toxic fragments [71]. Pro-
Tox produced better performance as compared with TOPKAT and
TEST, especially the sensitivity and precision (Table3). Moreover,Table 3
Performance of ProTox, TOPKAT, and TEST on the external set
Model ProTox (%) TOPKAT (%) TEST (%)
Sensitivity 73.1 44.8 46.3
Specificity 94.6 89.0 89.3
Precision 73.5 42.0 45.6
Coverage 91.8 89.4 78.6252 Jing Lu et al.