approximately 80,000 drugs and drug-like compounds [4]. Accelrys
Toxicity database collects six types of toxicological data for more
than 0.17 million compounds from the open scientific literature,
including acute toxicity, mutagenicity, tumorigenicity, skin and eye
irritation, reproductive effects, and multiple dose effects [5]. In
contrast, most of the open databases for chemical toxicities are
much smaller, such as TOXNET [6] and ISSCAN [7, 8].
It is estimated that the number of possible compounds has
already reached 10^60 [9]. This poses a large challenge for toxicity
evaluations of compounds using experimental screenings, which
need to use compound entities and multiple biochemical assays.
Therefore, particular interests have been raised to develop quick
and effective in silico models based on the information available for
recorded compounds, and supplement in vivo and in vitro testing
for predicting the toxicities of new compounds. For example, the
computational chemistry tools can use chemical structures and
experimental toxicities of the compounds as input data, calculate
molecular descriptors, and build prediction models using machine
learning methods. In this review, we discuss some illustrative mod-
els employing machine learning methods against a series of toxicity
end points, including acute toxicity, carcinogenicity, and inhibition
of the human ether-a-go-go-related gene ion channel (hERG).2 Acute Toxicity
Estimation of acute toxicity is one of the most common tasks in the
safety assessment of drug R&D, which represents the adverse
changes occurring immediately or within a short time after a single
dose of a compound or multiple doses given within 24 h
[10]. Median lethal dose (LD 50 ) or median lethal concentration
(LC 50 ) is common criterion for evaluating acute toxicity of com-
pounds in multiple species. The US EPA defined the toxicity cate-
gories based on LD 50 or LC 50 in 2014 (Table 1)[11]. The
compounds in Category I are considered highly toxic and fatal if
swallowed or inhaled. Category II means moderately toxic, and
Category III indicates slightly toxic, while Category IV is practi-
cally nontoxic [12]. Due to ethical reasons for avoiding the use of
animals, the alternatives, such as in silico models, are strongly
recommended by FDA, NIH, and EMEA [13–16]. Some pub-
lished models built by machine learning methods for acute toxicity
are discussed in detail below.2.1 Quantitative
Structure-Toxicity
Relationship (QSTR)
Models for Acute
Toxicity
For congeneric compounds with the same scaffold or mechanisms,
most of the relationship between the toxicity values and structural
descriptors approximates linear, and simple machine learning mod-
els, such as multiple linear regression (MLR) [17–19] and partial
least squares regression (PLSR) [20, 21], can generally achieve
good performance [22]. Furthermore, such linear regression248 Jing Lu et al.