Biomedical Engineering Reference
In-Depth Information
3
1
Experiments
Virtual
compounds
QSAR
2
FIGURE 10.5 Workflow of a QSAR paradigm. (1) Data collected from experiments and descriptors calcu-
lated from the structure of nanomaterials are combined to generate a QSAR model. (2) The model is then used
to predict the activity/toxicity of untested compounds that lie within the applicability domain of the model. (3)
Computed toxicities are used to streamline and prioritize new experiments. (Reprinted with permission from
Burello, E. and Worth, A. 2011. Nat. Nanotechnol. 6, 138-139.)
organisms, since the only descriptor in this prediction model is the enthalpy of formation of a gas-
eous cation with the same oxidation state as that in the tested metal oxide structure.
A report from Sayes and Ivanov (2010) claimed that a model has been set up using six different
physicochemical properties/features (engineered size, size in water, size in phosphate-buffered saline
(PBS) solution, size in cell culture medium, particle concentration, and zeta potential) for both TiO 2
and ZnO nanoparticles to predict cellular membrane damage (lactate dehydrogenase release). Two
commonly used pattern recognition mathematical models, multivariate linear regression, and linear
discriminant analyses (LDA) classification, were selected in designing predictive models. Taking
TiO 2 data from this report as an example, if the data are limited to the five TiO 2 features listed in
Table 10.3, linear regression cannot provide the correct framework to model cell membrane damage.
Therefore, the LDA classification was performed and its results showed that there was a significant
discrimination between a dense cellular membrane and a leaky cellular membrane. Table 10.3 is
the measurement of the resubstitution error based on the LDA classification. Figure 10.6 shows that
the combination of the nanomaterial's zeta potential with its engineered size and concentration in
ultrapure water produced a classifier with 0 resubstitution error. The multiple linear regression model
y = β 0 + β 1 x 0 + β 2 x 4 + β 3 x 5 was tested to two different data sets. The coefficient of determination for
the first set, which is composed of the entire available data, was r 2 = 0.70, while the same model pro-
duces r 2 = 0.77 when fit to the second set, which includes only dense- and disrupted-cell membranes
(Table 10.4). Although additional data are required for a thorough testing of existing hypotheses, the
possibility of using different QSARs models for different categories of biological responses cannot
be ignored. The authors also mentioned that the evaluation using additional features, such as water
solubility, bioavailable metals, and time-course evaluations, should be planned in the future.
10.5.3 g loBal M odels versus l local M odels
Defined by Cronin et al. (2009), global models are developed from data for large numbers of com-
pounds, crossing broad structural classes, and often mechanisms and modes of action, while local
models are more likely to be built on smaller numbers of compounds, often with some element of
structural and/or mechanistic similarity to them. Global models, definitely, cover broad chemical
space, but may include a number of descriptors, which do not have a direct, physicochemical sig-
nificance and may be formed with nonlinear techniques. There are many global models for toxic-
ity, such as the TOPKAT, M-CASE, and CAESAR models. In contrast, local models restrict the
domain of the model, but their advantages are that they may simply involve read-across or a simple
linear technique and are more accurate. There are various types of read-across and QSARs can be
regarded as sources of local models.
Search WWH ::




Custom Search