Abstract
The unique properties of nanomaterials, which mainly emanate from the small sizes and associated volume to surface ratio, results into various probable applications in different fields of industry. The same properties that favour potential applications may also pose environmental and health risks that result from unintentional exposure and environmental release. There are different categories of nanomaterials such as metal oxides and carbon-based nanoparticles. Within each class, nanomaterials come in different morphologies, chemical composition and surface modification.
With the ever-increasing production of nanomaterials, it is clearly not feasible to evaluate comprehensively the toxicity of these materials case by case. Therefore, more work to address nanomaterial’s toxicity towards human should be projected towards predictive models that are sensitive to small changes in the nanomaterial’s physicochemical properties and toxicity properties. This work, therefore, addresses toxicity of metal oxides nanoparticles (NPs) properties, that are computationally calculated with respect to exposed facets. This is achieved with the aid of using nanoQSAR models.
Metal oxide crystal structures were extracted from materials project website (https://materialsproject.org), from which, the surfaces ({100}, {110}, and {111}) were terminated from their optimized bulk structures. Recently advances have been directed towards growing NPs with specific facets with excellent efficiency, for unique applications. This is the basis for using computational methods in studying how the toxicity of NPs is influenced by their exposed facets.
Computations from Density Functional Theory (DFT) were employed to calculate molecular descriptors included HOMO (highest occupied molecular orbitals), LUMO (lowest unoccupied energy orbitals), band gap, hardness, chemical potential, enthalpy of formation, binding energy, Fermi energy, and electronegativity. These obtained properties were all with respect to the exposed low index facets. Periodic table properties were also used to compute descriptors for the classification model’s development. Furthermore, properties extracted from Transmission electron microscope (TEM) images using the NanoXtract software, were used as descriptors in the development of a classification model. Experimental properties, such as core
xii
and hydrodynamic size, were used as descriptors in the developed models. Toxicity data for 26 metal oxides was extracted from literature.
No gaps were found when the dataset's completeness was examined. The dataset was first put through a low-variance filter to eliminate any descriptors that did not exhibit considerable variance and could not improve the predictive ability of the model. The range of each descriptor was then made to follow a Gaussian distribution by applying the Z-score normalisation to account for the various numerical ranges. Using the BestFirst evaluator in conjunction with the Correlation-based Feature Selection technique, the descriptors utilized in the model's development were identified.
The Isalos Analytics Platform, which is powered by the Enalos + Tools, was utilized for model construction. On this platform, the cytotoxicity classification model was developed. The k-nearest neighbor algorithm's classification mode (EnaloskNN), an Enalos Chem/Nanoinformatics implementation, was used to create the metal oxide cytotoxicity model that is being presented. The EnaloskNN method is a lazy, instance-based method that computes the distance between the predicted endpoint and its k (k = 1, 2, 3...) nearest neighbors in the feature space Rn formed by the "n" identified significant descriptors. The model's best performance is used to define "k," which is used to make the prediction. The chemical potential, core size, electronegativity count, and enthalpy of formation for metal oxides were the four descriptors that were found to be significant in the model. To provide light on the relative toxicities of the nanoparticle, the relationship between these characteristics and the facets of the metal oxide is explored. The model and the supporting dataset can be accessed without charge on the NanoPharos database and the NanoSolveIT project cloud platform, respectively.
The k Nearest Neighbor (kNN) and Multilinear Regression (MLR) models were also developed using the IBM SPSS platform, producing both classification and predictive models. These models employed descriptors derived from electronegativity values calculated with respect to exposed facets. Significant descriptors were found to be facet electronegativity, effective mass, polaron effect, enthalpy of formation, work function and hydrodynamic size. A hybrid descriptor was also formulated in used in the model development. A comparison of generally metal oxide electronegativity and electronegativity derived from exposed facets was carried out to observe the impact
xiii
of using one over the other. Facet electronegativity was found to be more effective in improving the predictive power of the model.
The developed classification models exhibited a high level of predictability, with Ac values of 0.929, Sn values of 0.889, and Sp values of 1.000. The APD is 1.951 and Cohen's κ was calculated to be 0.851. A good forecast for both classes was indicated by the created model's MCC value of 0.861. The calculations were done for the k = 4 closest neighbours
Future studies can explore properties related to exposed facets and how the influence toxicity. Moreover, functionalised metal oxide, with know toxicity data can be considered in the development of the model. As such, more experimental data is required to improve the existing nano-QSAR models