Machine-learning Assisted Insights into Cytotoxicity of Zinc Oxide Nanoparticles

Zinc oxide nanoparticles (ZnO NPs) are commercially used as an active ingredient or a color additive in foods, pharmaceuticals, sun protection lotions, and cosmetic products. While the use of ZnO NPs in everyday products has not been linked to any serious health issues so far, the scientific evidence generated for their safety is not conclusive and, in most cases, could not be validated further in in vivo settings. To settle controversies arising from inconsistent in vitro findings in previous research focusing on the toxicity ZnO NPs, we combined the results of 25+ independent studies. One way analysis of variance (ANOVA) and classification and regression tree (CART) algorithm were used to pinpoint intrinsic and extrinsic factors influencing cytotoxic potential of ZnO in nanoscale. Particle size was found to have the most significant impact on the cytotoxic potential of ZnO NPs, with 10 nm identified as a critical diameter below which cytotoxic effects were elevated. As expected, strong cell type-, exposure duration- and dose-dependency were observed in cytotoxic response of ZnO NPs, highlighting the importance of assay optimization for each cytotoxicity screening. Our findings also suggested that ≥12 hours exposure to NPs resulted in cytotoxic responses irrespective of the concentration. Considering the cumulative nature of research processes where advances are made through subsequent investigations over time, such meta-analytical approaches are critical to maximizing the use of accumulated data in nano-safety research.


Introduction
Nanoscience deals with the phenomena that occurs in the nanometer range which is one billionth of a meter.While the conceptual roots of nanoscience were planted in the late 1950s, it was not until early 1990s that nanotechnology advanced enough to design structures, devices and systems at atomic and molecular scales (1).Nanoscale science and engineering is interdisciplinary in nature, requiring teams of researchers with different scientific backgrounds (e.g., physics, chemists, biologists, material scientists and engineers) working together to come up with new innovations and solutions to today's complex issues.The application of nanotechnology can span across different disciplines and research areas.Today, nanotechnology is explored in almost all existing domains ranging from high-strength materials and nanoscale sensors to electronic and opto-electronic devices (2).In parallel, novel properties of nano-scale materials are enabling new commercial markets such as next generation batteries and intelligent drug delivery systems (3,4).
Nanoparticles are commonly classified according to their origin (engineered or natural), dimensionality (0D, 1D, 2D or 3D), morphology (low or high aspect ratio), state (well-dispersed, aggregated etc.) or chemical composition (ceramic, polymeric, carbon-based or metallic) (5).Among different metal-based nanoparticles (NPs), zinc oxides (ZnO) stand out for their high UV-absorption capacity and solubility.They are commercially used as a bulking agent, filler or pigment in glass and ceramic products, foods, pharmaceuticals, sun protection lotions, and cosmetics (6).One of the early uses of ZnO NPs was in sunscreens due to their intrinsic UV absorbing properties and transparent nature (7).The use of nanosized ZnO (and also titanium dioxide) as an effective ingredient in modern sunscreens has created a long-lasting debate over their safety (8,9).In early 2010s (and onwards), both the regulatory bodies and the public have become increasingly aware of the potential threat posed by sunscreens formulated with nano-ingredients.The early findings related to potential hazards of ZnO NPs were mostly inconsistent, making it impossible to conclude with high certainty that nano-sized ZnO is ultimately safe to use in skin-contacting products (9).In the following years, it became clear that not all ZnO NPs should be treated the same from safety perspectives because physicochemical characteristics greatly affect cellular interactions and safety profiles of NPs (10).
Determining the potential harmful effects of NPs is critical to ensure that they are safe for human use.One effect of NPs that must primally be assessed is its cytotoxic potential, together with the factors contributing to their cytotoxicity (11).After two decades of research and detailed investigations, there is still no consensus on the main physicochemical properties driving cytotoxicity of NPs (12)(13)(14).In addition to intrinsic as-received properties of NPs and media-dependent surface characteristics, test conditions such as cell type, exposure concentration and duration have direct influence on the results of cytotoxicity assays.Figure 1 shows material-and assay-related parameters influencing different dimensions of NPs-protein and NPs-cell interactions.

Figure 1. Key parameters affecting the toxicity of NPs
ZnO NPs differ from their bulk counterparts in that their inherent complexity and medium-dependent characteristics make it very difficult to study their cellular interactions and effects.Moreover, the experimental differences in nano-hazard screening are directly reflected in test results, potentially leading to interexperimental inconsistencies.The aim of this study is to integrate published evidence on the cytotoxicity of ZnO NPs and to critically appraise bodies of evidence in their entirety.

Literature Search and Data Extraction
A systematic literature search was undertaken using the PubMed scientific search engine between 2010 and 2022.The following three terms were used for the initial article search: "zinc oxide", "nanoparticle*" and "cytotoxic*".The search returned 594 peer-reviewed research papers that were manually filtered according to the following inclusion criteria: (i) the core of the studied NPs must be zinc (and not a composite material); (ii) in vitro cytotoxicity data must be available and accessible; (iii) particle size data must be available; (iv) the unit of exposure concentration must be convertible to µg/mL; and (v) untreated cell control must be available.A total of 543 data points for 40 different ZnO NPs from the remaining 26 independent studies were included in the analysis.

Data Cleaning and Pre-processing
Data normalization (i.e., changing the values to a standard scale) is often used prior to statistical analysis when comparing features with different units or ranges.First, the units of measure were unified to minimize variability between different studies.The numeric data records describing the concentration were divided into ten subgroups.The cleaned data were randomly divided into training (75%) and test sets (25%), each involving a similar fraction of toxic and nontoxic groups.

Descriptive Statistics
One-way analysis of variance (ANOVA) was used to determine how strongly each of the categorical parameters describing NP, cell line, or assay characteristics was related to cytotoxicity.The strength and direction of the relationship between pairs of continuous variables were measured by Pearson's correlation coefficients.A box plot was used to display the distribution and skewness of the cell viability data among different subcategories.Significance was reported at p < 0.05 and p < 0.001 levels.

Machine Learning
Classification and Regression Tree (CART) was applied to partition the pre-processed data using a series of binary decisions.The method was set to regression, as the endpoint was a numerical value (% cell viability).The rpart package in R version 4.2.0 was used to implement all CART analyses.Regression trees were pruned through a 10-fold cross-validation process to remove branches providing the least error reduction.Refer to more specialized publications for details on the CART algorithm (15)(16)(17).

Results and Discussion
Description of data included in analysis.After a systematic data search and excluding data points that did not meet the data inclusion criteria, a total of 543 data records from 26 independent studies remained for evaluation.Each data record corresponds to a cytotoxic evaluation of individual nanoparticle.Figure 2 summarizes the main characteristics of the collected data.

Figure 2. Dataset description
Effect Sizes and Heterogeneity.A series of one-way ANOVAs were conducted to assess the influence of NPs and assay parameters on cell viability (Table 1).As expected, a strong negative correlation was observed between exposure dose and cell viability (p<0.001), with concentrations ≥20 µg/mL killing at least half of the cells.Similarly, cytotoxic profiles were detected after >12h exposure to ZnO NPs, with shorter exposure durations not causing significant toxicity.ANOVA also revealed that coating surfaces of NPs with amphiphilic polymers or thiol-containing acids could elevate its cytotoxicity, while green synthesis could help reduce the cytotoxic potential of ZnO NPs.These results highlight the importance of intrinsic materials characteristics and extrinsic experimental conditions on NP-induced cytotoxicity.A Pearson's correlation was run to assess the relationship between numeric parameters (particle size, hydrodynamic size, zeta potential, and concentration) and cell viability (%).The Pearson correlation coefficient of -0.22 suggested that cell viability and exposure concentration were moderately correlated in the opposite direction (Table 2).There was a positive correlation between particle size measured by TEM/SEM or DLS and cell viability, with NPs of larger diameter inducing less potent cell death.Interestingly, no direct correlation between zeta potential values and cell viability was observed.Next, a box plot was constructed to show the distribution of cell viability among different exposure durations and doses (Figure 3).As expected, higher concentrations of NPs and longer exposure durations led to higher levels of cytotoxicity relative to untreated cell control.The effect of extended exposure on cell viability was more pronounced at higher exposure doses.Machine learning.To identify the influence of material characteristics and experimental factors on the cytotoxic potential of ZnO NPs, the CART recursive partitioning analysis was employed.The bestperforming regression tree (Figure 4) was selected based on both cross-validation results and simplicity.The zeta potential measurements were not included in the decision tree analysis due to a high number of missing values (57%).Cross-validation error minimized at a tree size of 5 branches.The bestperforming regression tree given in Figure 4 included concentration, exposure duration, cell morphology, and particle size.The variable importance order was as follows: concentration > particle size > cell type > exposure duration > assay type > coating.In line with previous studies (18,19), our analysis showed that the potency of ZnO NPs to induce cytotoxic response is particle size-dependent.In particular, the primary particle size of 10 nm was found to be critical below which elevated cytotoxicity was seen.As expected, a strong positive linear relationship was observed between exposure duration and cytotoxic response.The longer the duration that cells are exposed to ZnO NPs, the greater the cytotoxicity.Most ZnO NPs were cytotoxic after 12 hours' exposure, especially at relatively higher doses (>20 µg/mL).As previously reported by Cierech et al., significant changes in cell viability were observed with increasing concentrations of ZnO NPs (20).The identified relationships between exposure conditions and cell viability results are also very much in line with the earlier investigations in the field (21)(22)(23)(24).For example, Khan and co-workers evaluated the toxic effects of ZnO NPs at different concentrations and demonstrated the role of reactive oxygen species generation in NP-induced cytotoxicity and genotoxicity (25).In another study, NP-induced DNA damage and cytotoxicity were evident after 6h exposure to 20 µg/mL of ZnO NPs (26).Taken together, the accumulated evidence on the cytotoxic and genotoxic potential of ZnO NPs suggests that the safety of ZnO NPs should remain a critical concern for all parties involved, including regulators, academicians, and industrial people.

Conclusion
Having a complete understanding of nanotechnology-related environmental, health, and safety (nano-EHS) issues is critical for bringing nano-enabled materials, products, and technologies to the mass market and ensuring their sustainable commercial use.Combining the results of multiple studies is increasingly applied in nano-EHS to pre-screen different NPs based on their toxicity potential and to look for early signs of harmful effects (10,27,28).By statistically integrating findings of independent nanotoxicity-screening studies, it is possible to get a more accurate representation of complex behavior of nano-systems in cellular systems.In this study, we examined the factors contributing to the cytotoxic potential of ZnO NPs using meta-analytic data covering 543 data points from 26 independent studies.The main aim here was to underpin parameters that potentially control the biological effects of ZnO NPs.While the data used in this study was for ZnO NPs, similar models can be developed for different-core metallic NPs with small changes in the framework.Such data-driven models and the insights they provide are key for the simultaneous realization of safe-by-design and quality-by-design concepts that aim to ensure the safety and quality of NPs in an early stage of the innovation process (Figure 5).

Figure 3 .
Figure 3. Box plot of changes in cell viability (%) as a function of exposure concentration (dose) category, grouped by exposure duration.Circles outside the plot represent outliers beyond the 10th and 90th percentiles.

Figure 4 .
Figure 4.The best-performing regression tree predicting cell viability of ZnO NPs

Figure 5 .
Figure 5. Integration of safety and quality by design concepts through machine learning

Table 1 .
One-way ANOVA results

Table 2 .
Pearson correlation results