The various aspects of nanotechnology have been advancing for over half a century and continue to grow rapidly [1
]. An increased usage of nanomaterials results in their release into the environment, making it necessary to determine the hazardous effects of these materials [4
]. Determining the toxicity of materials is one of the main steps in safe materials design [5
]. The majority of current studies are focused on the investigation of the environmental safety of nano-sized materials and devices [6
Every year, an enormous amount of money is spent in both academia and industry on testing new hypotheses and the research and development of new nanomaterials [8
]. Various alternatives to animal testing have been proposed to overcome some of the drawbacks associated with animal experiments [9
]. The idea of the 3Rs approach (reduction, refinement, and replacement) is suitable for the investigation of nanoparticles (NPs). Different strategies have been proposed to reduce the number of experiments, from the standardization of testing conditions and extrapolation to the application of computational techniques [5
]. For instance, as an alternative to animal testing, zebrafish has been replaced by its most representative enzyme: the zebrafish hatching metalloprotease, ZHE1 [11
]. In the current project, we have developed a computational model based on the in vitro data of ZHE1 inhibition.
Over the years, different computational methods have been developed, but the quantitative structure–activity/properties relationships (QSAR/QSPR) approach is the most time-effective one [13
]. The fundamental assumption of QSAR modeling is that the features of molecular structure are responsible for physical or biological properties of the studied samples [16
]. Each compound could be represented by a set of different numbers (known as “descriptors”) [17
]. For this purpose, the features of each chemical structure are defined by different physical or mathematical concepts and then transformed into descriptors [17
]. In the last ten years, more than fifty QSAR/QSPR models for nanoparticles (known as “nano-QSAR models”) have been published [18
Unique properties of NPs are governed by different characteristics related to the specific organization of NP (release of ions from a surface, huge contact area, aggregation effects, etc.). Therefore, researchers usually utilize experimental properties as descriptors: size, volume concentration, solubility, zeta potential, Brunauer–Emmett–Teller (BET) measurements (shape, surface area), reactivity, etc. [20
]. Experimentally-based descriptors have a clear interpretation, which makes them a very valuable class of NP descriptors. However, experimental parameters can contain systematic or chance error, as the quality of measurements greatly depends on the equipment and the skills of the researcher [17
]. This makes these descriptors applicable only for representing co-measured data (data obtained in the same laboratory).
As many existing models successfully combine simple descriptors with experimental data, we suggested that a successful depiction of the properties or mechanisms of action would become possible by means of a hierarchical combination of descriptors. The main idea of the hierarchical approach is a simultaneous use of descriptors that reflect the NPs’ structure at different levels of the organization of molecules: from ions to aggregates [23
The objective of the current paper is the development of a theoretical 3R model that provides a safety assessment of the nanoparticles towards zebrafish. To reach this goal, a combination of theoretical descriptors with experimental parameters was applied for nano-QSAR modeling. As it has recently been pointed out, nanomaterials can demonstrate non-linear changes in toxicity. To address this problem in nano-QSAR modeling, a non-linear M5P tree approach was used. The general idea of the technique that was used is the creation of a conventional decision tree combined with linear regression functions [24
]. Additionally, causal relationships between the activity and the descriptors of the developed model were analyzed.
The developed nano-QSAR model has high statistical characteristics: R2adj = 0.85 and RMSE = 0.062 for the training set, R2adj = 0.69 and RMSE = 0.176 for the validation (bagging) set, R2adj = 0.80 and RMSE = 0.112 for the external validation (test set). Prediction for Cr2O3 in the training set was outside of the three-sigma deviation for the developed model (outlier).
The developed model to be discussed is non-linear. However, for the readers’ convenience, the observed–predicted diagram is presented in a classical way (in Figure 1
). The straight line represents perfect agreement between the experimental and the calculated values.
The developed model contains four descriptors: the hydrodynamic size (∅hyd
), density (ρ), the Wigner–Seitz radius (rwz
), and the covalent index (CI
). In Figure 2
, the distributions of values for the descriptors are represented using diagrams. QSAR methodology assumes that the structural variation causes the variation of physical or biological properties; therefore, adequate distribution of properties (descriptors) within the training/test set is crucial. As one can see from Figure 2
, the distributions of selected descriptors varied considerably.
The nano-QSAR model is presented in Figure 3
. The developed non-linear model contains three nodes. The first node (group 1) covered ~42% of the training set (eight NPs): Ni2
. Similarly, the second node (group 2) covered ~31.5% (six NPs): Fe2
, CoO, TiO2
. The third node (group 3) covered ~16% (three NPs): Cr2
, ZnO, CuO. The remaining nanoparticles (group 4, ~10.5%, two NPs) were not included in the three main nodes: CeO2
. The percentage of involvement for each descriptor for all nodes (relative importance, RI) is presented in Table 1
In order to evaluate the causality of the developed nano-QSAR model, a probability of causal relationships between toxicity and selected descriptors was measured. Direct and clear cause–effect relationships were found for the hydrodynamic size and toxicity with p = 0.05 in the first node (the Wigner–Seitz radius) and p < 0.001 for the remaining nodes.
The percentage of relative importance of a descriptor reflects its mean influence. One of the most important advantages of non-linear models (such as decision tree-based models) is their potential to simultaneously describe different mechanisms of action. Each individual node (“leaf”) in a tree could be applied only to a specific group of nanoparticles. It is very critical for a numerical representation of nanoparticles, since different types of nanoparticles can affect certain targets (enzymes, cells, etc.) in opposite ways. As was mentioned in the introduction, the mechanisms of a nanoparticle’s action could vary as solubility, aggregates formation, and surface charge change.
Based on the above, low overall relative influence (8.2%) for the covalent index does not mean that it has low impact on the nanoparticle’s action. Group 4 is driven by the CI
); therefore, the CI
reflects only one mechanism of action (100%). It is critical to note that nanoparticles that were grouped by the CI
, group 4)—namely ZnO and CuO in the training set and NiO in the validation set—demonstrated the highest ZHE1 inactivation.
The overall schematic representation of suggested toxicity mechanisms is outlined in Figure 4
. The toxicity of nanoparticles could be related to a two-step process. At the first step, ions are released from the nanoparticle’s surface through the dissolution mechanism (Figure 4
, Case 1). The likelihood of releasing depends on the environmental conditions and on the general trends of dissolution of the selected nanoparticle. Next, released metal ions interact with metal-sensitive binding sites in the active center of the zebrafish hatching enzyme, ZHE1 [11
ZHE1 is a zinc metalloprotease. Generally, different divalent metal ions can replace the zinc ion in the enzyme’s binding site. However, binding affinity differs among a series of metal ions. For instance, Holt et al. demonstrated that the binding affinity of metals to metallothionein changed in the order Zn2+
]. It was found that thermolysin activity increased when cobalt was present [26
]. Replacement of Zn2+
ion by Cu2+
in the catalytic center of astacin increased enzymatic activity; inclusion of Ni2+
inactivated the enzyme [27
]. Similar results were obtained recently: metal substitution of Zn2+
in dipeptidyl peptidase maintained catalytic activity [28
]. On the contrary, Cu2+
that substituted dipeptidyl peptidase could not restore the catalytic activity. These findings were supported by the analysis of coordination geometries: Cu2+
coordination geometry is very rigid, while the coordination geometries of Zn2+
, and Mn2+
are flexible [25
]. Moreover, even Zn2+
overdose can alter metalloprotease activity: Burkholderia cenocepacia
ZmpA zinc metalloprotease activity was lowered by an excess of Zn2+
As it was pointed out by Lin et al., the ionic radius of metal ions, associated ion-ligand coordination chemistry, and binding constants could determine substitution in the enzyme center [12
]. Based on the above, the theoretical descriptor CI
simultaneously reflects the properties of released metal ions (charge and electronegativity) and interactions with the metal-binding amino acids in the enzyme center. Using numerical analysis, polynomial regression for the covalent index and the enzymatic activity of ZnO, CuO, and NiO (R2
= 0.85) were identified. Therefore, the descriptor CI
represents an attractive theoretical parameter that reflects mechanisms of action of released metal ions.
As one can see from Table 2
, an influence of the hydrodynamic size and the Wigner–Seitz radius have the highest impacts and contribute equally to the enzyme inhibition. We suggest that the hydrodynamic size and the Wigner–Seitz radius represent similar—or possibly the same—mechanism of action (Figure 4
, Case 2 and Case 3). For instance, nanoparticles that were characterized by higher values of the Wigner–Seitz radius (rwz
> 0.196, combination of area 1 and area 2 in Figure 3
) demonstrated simple linear relationships between enzymatic activity and the hydrodynamic size (Pearson’s correlation coefficient r
With a high probability, the Wigner–Seitz radius indirectly reflects the fraction of molecules on the nanoparticle’s surface, which could be responsible for toxicity or inhibitory activity (Figure 4
, Case 2). In general, the high influence of nanoparticles is caused by their large specific surface area (Figure 4
, Case 2). Indeed, metal oxides with higher values of the Wigner–Seitz radius usually have a smaller surface-to-volume ratio (when sizes of studied nanoparticles are the same). A smaller surface-to-volume ratio leads to lower activity. This is in agreement with the developed model (see group 1 and group 2, in Figure 2
). The groups of nanoparticles with rwz
> 0.196 demonstrated lower activity.
Moreover, as it was demonstrated in our previous contribution, the “liquid drop” model (LDM)-based Wigner–Seitz radius (rw
) and mass density (ρ) are mutually-related descriptors [30
]. The Wigner–Seitz radius (rw
) and mass density (ρ) are not directly correlated to each other in statistical terms, but the relationship between ρ and rw
has a straightforward mechanistic interpretation and can be deducted from the basic LDM equation (Equation (1)).
As it was mentioned, the hydrodynamic size and the Wigner–Seitz radius are also related to the third possible mechanism (Figure 4
, Case 3). Another important process represents the self-assembly of nanoparticles in large aggregates or agglomerates in a culture medium (Figure 4
, Case 3). Large nanoaggregates are not able to reach the binding site of the enzyme. If these nanoaggregates are poorly soluble, they do not release substantial amounts of metal ions. As a result, metal ions are not capable of reaching the binding site of the enzyme. In general, poorly soluble or non-soluble nanoparticles could inactivate enzymes only blocking outer surfaces. In other words, it is possible that certain metal oxide nanoparticles adsorb the target enzyme. However, at the same time, larger aggregates reach enzymes (or cells) faster than smaller particles due to gravity sedimentation. For example, Ni2
, and Figure 4
, Case 3) were characterized by a large hydrodynamic diameter (665.8 and 831.7 nm, respectively) and demonstrated higher enzyme inhibition than smaller nanoparticles aggregated from group 1 (In2
, and La2
Summarizing the data presented in Figure 4
, one can conclude that the release of metal ions from the surface of NP could cause their interaction with metal-sensitive binding sites in the active center of the enzyme (Case 1, Figure 4
); the high influence of nanoparticles is caused by their large specific surface area (Case 2, Figure 4
); aggregation can positively or negatively influence the dynamics of metal ion accumulation (Case 3, Figure 4