Nano-QSAR modeling for predicting the cytotoxicity of metal oxide nanoparticles using novel descriptors

Abstract

Computational approaches have evolved as efficient alternatives to understand the adverse effects of nanoparticles on human health and the environment. The potential of using Quantitative Structure–Activity Relationship (QSAR) modeling to establish statistically significant models for predicting the cytotoxicity of various metal oxide (MeOx) nanoparticles (NPs) has been investigated. A novel kind of nanospecific theoretical descriptor was proposed by integrating codes of certain physicochemical features into SMILES-based optimal descriptors to characterize the nanostructure information of NPs. The new descriptors were then applied to model MeOx NP cytotoxicity to both Escherichia coli bacteria and HaCaT cells for comparison purposes. The effects of size variation on the cytotoxicity to both types of cells were also investigated. The four resulting QSAR models were then rigorously validated, and extensively compared to other previously published models. The results demonstrated the robustness, validity and predictivity of these models. Predominant nanostructure factors responsible for MeOx NP cytotoxicity were identified through model interpretation. The results verified different mechanisms of nanotoxicity for these two types of cells. The proposed models can be expected to reliably predict the cytotoxicity of novel NPs solely from the newly developed descriptors, and provide guidance for prioritizing the design and manufacture of safer nanomaterials with desired properties.

---

1 Introduction

Nanoparticles (NPs), which are defined as microscopic particles having at least one dimension <100 nm,^1,2 have gained widespread applications in different areas due to their unique and beneficial physicochemical properties, such as small size, large surface-area-to-volume ratio, porosity, surface charge, and infinite possibilities for modifying their surface chemistry.^3–5 For instance, NPs are widely used in cosmetics and sunscreens,⁶ medicine and biology (including protein detection,⁷ DNA structure probing,⁸ phototherapy agents,⁹ imaging tools,¹⁰ gene delivery carriers,^11,12 and drug delivery systems^13,14), etc. However, their extraordinary properties may also give NPs new but harmful biological activities, such as the toxic effects of NPs on humans and the environment.^3,15–18

The evaluation of the toxic effects was usually performed using experimental in vivo toxicological studies and in vitro short-term cell-based assays. However, toxicological tests are costly, laborious, time-consuming and sometimes unethical, making it impractical to test all the NPs-based products available. Therefore, there is an imperative need to develop efficient computational in silico methods, such as Quantitative Structure–Activity Relationship (QSAR), to screen and predict the toxicity of NPs in a rapid and cost-efficient way prior to their synthesis.

QSAR is a mathematical method that relates the properties of interest to the molecular structures of chemicals which are represented by a variety of molecular descriptors. This computational method is based on the assumption that the variation in the properties or biological activities of a chemical is determined by the changes in its molecular structure. QSAR can be expected to capture the complex relationships between the molecular structures (microscopic) and properties or activities (macroscopic) of chemicals without requiring detailed knowledge of the mechanisms of interaction. Also, it is time- and money-saving, and could reduce the number of animal testing as well as reveal the toxicity mechanism. Consequently, QSAR is widely applied in drug discovery and physicochemical properties modeling, which have been extensively reviewed elsewhere.^19–22 Its growing importance is also accepted by regulators for the assessment and management of new industrial chemicals, such as the European REACH as well as the Australian NICNAS.

Recently, more and more investigations revealed that there is a strong need to extend the traditional QSAR paradigm to NPs and to develop “nano-QSAR” models for predicting the activity/property of newly synthesized NPs as well as guiding the experimental design of new NPs with improved activity.^2,23–25 In the report on nanosafety published in 2013 by the European Commission, development of appropriate nano-QSAR models to identify high concern NPs and predict relevant endpoints of toxicity and ecotoxicity had been considered to be of prime importance for NPs hazard assessment. However, although QSAR has achieved substantial progress in recent years for modeling the traditional chemicals, its application to NPs is still limited. The reasons may be mainly attributed to NPs' more complex and less amenable features to “nano” structure characterization, endpoint description, and physical interpretation than those for traditional chemicals. Particularly, because the exact composition of a given NP is usually not known and some highly complex atoms may be included in large three-dimensional nano-structures, the classical molecular descriptors are unable to express specificity of “nano” structures, leading to the lack of sufficient molecular descriptors appropriate for nano-QSAR modeling.

Recently, a few efforts have been made to develop nano-QSAR models to correlate nano-structures with activities of NPs, such as cytotoxicity,^26–30 cellular uptake,^31–35 and smooth cell apoptosis.³⁵ The nano-structure characteristics of NPs can be usually described by either experimentally measured physicochemical properties (i.e. size, shape, surface area, zeta potential, overall charge) or theoretically calculated molecular descriptors (i.e. quantum-mechanical descriptors, periodic table-based descriptors, SMILES based descriptors). However, it should be mentioned that in most cases the determination of the experimental descriptors itself is always laborious and time-consuming, thus the data of these descriptors are always unknown, especially for the NPs which are newly developed or even not yet synthesized. As for the most employed quantum-mechanical descriptors, certain quantum-chemical background is required to be able to perform the calculations, and the descriptors are a little bit computationally demanding for such large NP systems comprising hundreds of atoms. Besides, most existing studies have been carried out on modeling toxicity of NPs to prokaryotic systems (i.e. Escherichia coli bacteria, E. coli), while very few studies have focused on toxicity to eukaryotic systems. Information is only limited on the modeling of toxicity of metal oxide (MeOx) NPs to a human keratinocyte (HaCaT) cell line for dermal exposure.^28,30

In this study, we aim to present new promising nano-specific descriptors to develop reliable, predictive and freely-accessible nano-QSAR models for predicting the cytotoxicity of MeOx NPs. The new descriptors were improved and developed from the traditional SMILES-based optimal descriptors, and capable of characterizing the structural and chemical properties of MeOx NPs. The new descriptors were applied to model MeOx NPs' cytotoxicity to both E. coli bacteria (prokaryotic cells) and HaCaT cells (eukaryotic cells) for comparison purpose, leading to models with high statistical quality as well as acceptable interpretability. Moreover, the effect of size variation on cytotoxicity of MeOx NPs to both cells were considered and investigated here, since most of the existing researches always nullify a typical problem that many activities or properties of NPs change for different size ranges. After implementing extensive validation using multiple strategies, the obtained QSAR models are expected to provide critical support for predicting the toxicological effects of novel NPs solely from their nano structures, as well as designing and manufacturing safer NPs with desired properties.

2 Materials and methods

2.1 Dataset

MeOx nanoparticles are one of the most important classes of engineered nanomaterials, which are used mainly in microelectronics, textiles, cosmetics and paints. In this study, two different datasets of toxicity of MeOx NPs to both E. coli and HaCaT cells were employed for the nano-QSAR modeling. All the experimental toxicity data in both datasets were taken from the previous publication for comparison purposes and presented in the ESI.†

Case study 1: cytotoxicity in E. coli bacteria of MeOx NPs. The dataset contains 17 different MeOx based NPs with sizes ranging from 15 to 150 nm.²⁶ Their cytotoxicity in E. coli bacteria are expressed as logarithms of the inverse molar effective concentration (log(1/EC₅₀)), which varied from 1.74 to 3.51 mol L⁻¹.

Case study 2: cytotoxicity in HaCaT cells of MeOx NPs. The dataset contains 18 different MeOx NPs with sizes between 15 and 150 nm.³⁰ Their cytotoxicity in HaCaT cells are expressed as logarithms of the inverse molar median lethal concentration (log(1/LC₅₀)), which varied from 1.76 to 3.32 mol L⁻¹.

2.2 Descriptors development and calculation

The structural complexity and dynamic behavior of NPs always complicates the nano-structure characterization and nano-descriptor calculations. In the case of the widely used quantum-chemical descriptors, the calculations for NPs with sizes between 15 and 150 nm (as in this study) were time-consuming or even not feasible, despite the previous studies have always maximally simplified the structural models into smaller fragments (clusters) for all involved NPs. Moreover, their accuracy is often disputable. Thus, the extensive quantum-mechanical calculations for relatively large molecular clusters might be avoided. More recently, in order to simplify the representation of nano-structure characteristics and possible interactions at the nano-level without quantum-chemical modeling, the simplified molecular input-line entry system (SMILES) based descriptors are gradually becoming an attractive alternative to “classic” descriptors and have been successfully employed to correlate with various endpoints of NPs.^27,34 The CORAL software (http://www.insilico.eu/coral) is a sophisticated provider of SMILES-based optimal molecular descriptors, which could be freely-accessible online.

However, after a critical analysis of the results obtained from the developed nano-QSAR models, some drawbacks of the used SMILES-based descriptors can be found for describing the endpoints for some NPs. For example, some different MeOx NPs may contain the same descriptor values and thus lead to the same predicted endpoint values for the developed models, although their experimental endpoint values might be quite different indeed. Moreover, interpretability of nano-QSAR models developed from SMILES based descriptors is limited, since such descriptors are obtained solely from SMILES notations. Consequently, to overcome the defect of traditional SMILES-based optimal descriptors, some additional features could be considered and included into the descriptors to predict and interpret the endpoint of NPs more accurately.

Hence in this study, a novel kind of nano-structure descriptor is proposed to build up predictive models for cytotoxicty of MeOx NPs. Some basic physicochemical features, such as molecular weight, cationic charge, and mass percentage of metal elements, were taken into account and included into the SMILES based optimal descriptors to improve the characterization of nano-structure information. That is to say, MeOx NPs can be represented by SMILES-based optimal descriptors integrating codes of certain physicochemical features. We named the corresponding new descriptors as Improved SMILES-Based Optimal Descriptors, which are calculated as following:

DCW(threshold, N_epoch) = ∑ CW(S_k) + ∑ CW(SS_k) + ∑ CW(C_k) + ∑ CW(CC_k)

where S_k and SS_k are one- and two-element SMILES attributes obtained from the SMILES; CW(S_k) and CW(SS_k) are the correlation weights of the attributes. Similarly, the C_k and CC_k are codes of k-th physicochemical features; CW(C_k) and CW(CC_k) are the correlation weights for features.

The SMILES can be obtained by common software like ChemSketch, and the codes of features can be defined by the following scheme:^36,37

(I) Standardization of feature X_k based on the formula

(II) Distinction of physicochemical features in light of scale (Fig. 1) into one of classes from 1 to 9. Liters of A, B, C represented molecular weight, cationic charge, and mass percentage of metal elements, respectively.

Fig. 1 Distinction of standardized physicochemical features into classes 1 to 9 according to its value.

The construction of S_k (or C_k) and SS_k (or CC_k) could be represented as the following:

ABCDE → A, B, C, D, E

ABCDE → AB, BC, CD, DE

The optimal descriptors are then calculated with the correlation weights of active physicochemical features by the Monte Carlo optimization. The threshold is coefficient for classification of features into two classes (noise and active). The N_epoch is the number of epochs of the Monte Carlo optimization.

The optimization is the following:³⁷ for each attribute, CW is determined initially by setting the start values of all CWs to 1 ± 01* random. The random is generator of random value of range (0, 1). The regular order of number of attributes (i.e., 1, 2, 3, 4, 5,…) is replaced by a random sequence (e.g., 3, 1, 5, 2, 4,…).

A starting correlation coefficient (R₁) between endpoint and descriptor on the training set is calculated. In a generated random sequence, each attribute correlation weight CW_i was modified with the algorithm.

(1) ΔCW_i := 0.001 × CW_i; Eps := 0.01 × ΔCW_i;

(2) CW_i := CW_i + ΔCW_i;

(3) Calculation of R₂ (the correlation coefficient after modifying CW_i);

(4) If R₂ > R₁, then R₁ := R₂; go to step 2;

(5) CW_i := CW_i − ΔCW_i;

(6) ΔCW_i := −0.5 × ΔCW_i

(7) If absolute value (ΔCW_i) > Eps then go to step 2.

The steps of 1–7 are carried out for all CWs, and one can repeat this algorithm from point of generation of random sequence. If the increasing of correlation coefficient becomes less than 0.001, the process will stop.

Data on these features of various MeOx NPs can be easily obtained from molecular formula and information acquired from periodic table, which are presented in the ESI.†

Moreover, since size has been proven to play an important role for both bioactivity of NPs³⁸ and their toxicity,³⁹ understanding its effect on activities of NPs is necessary to design NPs with desired features. Interestingly, the widely used quantum-chemical descriptors for nanotoxicity modeling were always calculated based on the assumption that the simplified clusters must be of the same size. Moreover, most previous studies^40–42 revealed that the most significant size-dependent changes of NPs' properties occur below about 5 nm, while the particle size did not affect activity for NPs with sizes between 15 and 90 nm, since all nanopowders resulted in similarly sized aggregated particles in water suspension, regardless of the powder size.⁴³ Therefore, the property changes for NPs with sizes between 15 and 90 nm had always been neglected.

In this study, in order to further verify the size effects of MeOx NPs on their cytotoxicity, we also included the individual sizes and aggregation sizes as the codes of physicochemical features into the new improved descriptors to develop corresponding QSAR models for comparison purpose. Experimental data on the individual sizes and aggregation sizes of various MeOx NPs are taken from the literature²⁸ and presented in the ESI.†

In this way, two different forms of the new improved SMILES-based optimal descriptors (namely size-independent and size-dependent) were employed to model the cytotoxicity to both E. coli bacteria and HaCaT cells of diverse MeOx NPs, respectively.

2.3 Dataset splitting

The dataset splitting plays an important role in the development of statistically significant and reliable nano-QSAR models. Here the toxicity dataset for E. coli was split into the training and test sets according to the division of the Puzyn's work.²⁶ Then the dataset for HaCaT cells was split according to the following principles: (i) the split is random keeping the highest and lowest toxic NPs in the training set; (ii) the test set NPs should lie within the chemical space occupied by the training set NPs and cover all types of oxides (MeO, MeO₂, Me₂O₃). The division details of NPs for both cells were shown in Tables 1 and 2, respectively. The training set was then utilized for model development, while the test set was used to assess the predictivity of the models.

Table 1 Experimental and predicted log(1/EC₅₀) values of MeOx NPs to E. coli

MeOx	Experimental values	Model 1		Model 2		Set
		DCW	Predicted values	DCW	Predicted values
Al2O3	2.49	9.284	2.500	17.354	2.456	Training
Bi2O3	2.82	11.128	2.990	20.833	2.950	Training
CuO	3.2	12.316	3.306	—	—	Training
Fe2O3	2.29	9.671	2.603	16.725	2.367	Training
In2O3	2.81	9.877	2.658	19.201	2.718	Training
SiO2	2.2	7.160	1.935	15.263	2.159	Training
SnO2	2.01	7.754	2.093	14.236	2.014	Training
TiO2	1.74	7.013	1.896	12.475	1.764	Training
Y2O3	2.87	9.877	2.658	19.831	2.808	Training
ZnO	3.45	12.316	3.306	24.314	3.444	Training
CoO	3.51	12.376	3.322	25.659	3.635	Test
Cr2O3	2.51	9.796	2.636	17.984	2.546	Test
La2O3	2.87	10.752	2.889	18.485	2.617	Test
NiO	3.45	12.974	3.481	23.054	3.265	Test
Sb2O3	2.64	8.738	2.354	19.201	2.718	Test
V2O3	3.14	9.695	2.609	19.330	2.737	Test
ZrO2	2.15	7.160	1.935	13.735	1.942	Test

---

Table 2 Experimental and predicted log(1/LC₅₀) values of MeOx NPs to HaCaT cells

MeOx	Experimental values	Model 3		Model 4		Set
		DCW	Predicted values	DCW	Predicted values
Al2O3	1.85	20.318	1.818	23.842	1.856	Training
Bi2O3	2.5	26.579	2.468	32.010	2.491	Training
Cr2O3	2.3	24.772	2.280	29.555	2.300	Training
In2O3	2.92	31.196	2.947	37.523	2.919	Training
La2O3	2.87	30.570	2.882	36.861	2.868	Training
NiO	2.49	28.849	2.704	31.841	2.478	Training
Sb2O3	2.31	25.055	2.310	29.764	2.316	Training
SnO2	2.67	28.619	2.680	34.274	2.667	Training
V2O3	2.24	24.203	2.221	28.698	2.233	Training
WO3	2.56	27.409	2.554	33.026	2.570	Training
ZnO	3.32	33.342	3.170	42.750	3.326	Training
ZrO2	2.02	22.159	2.009	26.010	2.024	Training
Mn2O3	2.64	28.312	2.648	—	—	Training
CoO	2.83	25.846	2.392	30.912	2.406	Test
Fe2O3	2.05	21.794	1.971	24.791	1.930	Test
SiO2	2.12	21.056	1.895	24.716	1.924	Test
TiO2	1.76	21.282	1.918	22.564	1.756	Test
Y2O3	2.21	21.635	1.955	25.632	1.995	Test

---

2.4 Model validation

Model validation is of crucial importance to nano-QSAR models. Here all the developed models were evaluated using various OECD recommended criteria proposed in classic QSAR literatures.⁴⁴

Both squared correlation coefficient for fitting (R²) and Root-Mean-Square Error (RMSE) were employed to indicate the model calibration ability, while both the internal and external validations were employed for predictive capability. Here the most reliable internal validation method of leave-many-out (LMO, 20% out) cross-validation (Q_LMO²) was employed to indicate the model robustness and internal predictive ability. The external validation is a significant and necessary validation method used to determine both the generalizability and true predictive capability of the QSAR models for new NPs. Here a separate test sub-set of each dataset which was kept out during the training process was used for external validation. Moreover, in order to avoid obtaining the QSAR results by chance or obtaining non-general conclusions, all the developed models were tested on a sufficiently large number of compounds (20% of the dataset) in the test sub-sets.⁴⁵ The developed models were also subjected to Y-randomization test to check the model reliability and robustness.^46–49

3 Results and discussion

3.1 Case study 1: nano-QSAR modeling of cytotoxicity to E. coli

Here two different forms of the improved SMILES-based optimal descriptors (considering the size effect or not) were employed to model the cytotoxicity to E. coli, respectively. The corresponding nano-QSAR models were obtained and presented as following:

Model 1.

log(1/EC₅₀) = 0.0321 (±0.1443) + 0.2658 (±0.0141) × DCW(6,11), n = 10, R² = 0.8891, Q_LMO² = 0.8378, s = 0.179, F = 164, p < 0.0001

Model 2.

log(1/EC₅₀) = −0.0076 (±0.0306) + 0.1420 (±0.0020) × DCW(6,17), n = 9, R² = 0.9824, Q_LMO² = 0.9745, s = 0.007, F = 391, p < 0.0001

In both models, n is the number of NPs in training set; Q_LMO² is cross-validated R²; s is standard error; F is Fischer ratio; p is p-value.

It should be mentioned that there were no individual and aggregation sizes information available in the literature for CuO NP. Thus the original dataset for model 1 was reduced to 16 when developing model 2 for evaluating the size effects of NPs on their cytotoxicity.

Both models were then utilized to predict the log(1/EC₅₀) values of the 7 NPs in the test set for external validation. The predicted values were presented in Table 1, while the main statistical parameters were shown in Table 3. One can see that both models are statistically satisfactory, with the RMSE values for both sub-sets as low as possible. Plots of the predicted log(1/EC₅₀) values against the experimental ones for both models were shown in Fig. 2, which indicated a reasonable agreement between the predicted and experimental values across the entire dataset.

Table 3 Comparisons of statistical parameters between the presented and previous models

Endpoints	Models	Training set			Test set
		R^2	RMSE	n	R^2	RMSE	n
E. coli	Puzyn et al.^26	0.85	—	10	0.83	0.19	7
	Sizochenko et al.^28	0.93	0.13	13	0.78	0.32	3
	Model 1	0.8891	0.181	10	0.8181	0.257	7
	Model 2	0.9824	0.065	9	0.8670	0.216	7
HaCaT cells	Gajewicz et al.^30	0.93	0.12	10	0.83	0.13	8
	Sizochenko et al.^28	0.96	0.10	14	0.92	0.12	3
	Model 3	0.9606	0.075	13	0.8281	0.250	5
	Model 4	0.9997	0.007	12	0.9905	0.206	5

---

Fig. 2 Comparisons between the predicted and experimental log(1/EC₅₀) values of toxicity to E. coli.

Moreover, it is noteworthy that the results obtained by model 2 are better than those by model 1, for both the training and test sets, as the R² and RMSE values have been obviously improved. That is to say, comparing to model 1 without considering the size effect, model 2 considering the size effect here can offer a better predictive ability.

3.2 Case study 2: nano-QSAR modeling of cytotoxicity to HaCaT cells

The two different forms of the new proposed descriptors were then employed to model the cytotoxicity to HaCaT cells. The corresponding nano-QSAR models were presented as follows:

Model 3.

log(1/LC₅₀) = −0.2909 (±0.0664) + 0.1038 (±0.0027) × DCW(1,3), n = 13, R² = 0.9606, Q_LMO² = 0.9393, s = 0.008, F = 268, p < 0.0001

Model 4.

log(1/LC₅₀) = 0.0012 (±0.0048) + 0.0778 (±0.0001) × DCW(1,3), n = 12, R² = 0.9997, Q_LMO² = 0.9996, s = 0.007, F = 1273, p < 0.0001

There were also no exact individual and aggregation sizes information available in the literature for Mn₂O₃ NP. Thus the original dataset for model 3 was reduced to 17 when developing model 4.

The predicted log(1/LC₅₀) values for both models are presented in Table 2. Both models are statistically satisfactory with the RMSE values for both sub-sets as low as possible. Plots of the predicted values against the experimental ones for both models were shown in Fig. 3.

Fig. 3 Comparisons between the predicted and experimental log(1/LC₅₀) values of toxicity to HaCaT cells.

Also, it should be noticed that the performance of model 4 are obviously superior to those of model 3, especially for the test set, although model 3 also satisfies the minimum criteria for a successful QSAR model (R² > 0.6, Q_LMO² > 0.5).⁵⁰ This strongly suggested that the particle sizes had a significant effect on the cytotoxicity of MeOx NPs to HaCaT cells.

3.3 Model stability validation and results analysis

All the developed models were then tested for chance correlation to further analyze the model stability. The Y-randomization test was performed on the dataset for 10 times for each model. The obtained R² values of the new generated models were presented in Table 4. As expected, all the models generated had produced low R² values, which were much lower than the ones calculated when the dependent variables were not scrambled. It indicated that only the correct dependent variables can be used to generate reasonable models, and the chance correlation had little or even no effect in the presented models.

Table 4 Results of Y-randomization tests

Probe of Y-scrambling	E. coli				HaCaT cells
	Model 1		Model 2		Model 3		Model 4
	Training set	Test set	Training set	Test set	Training set	Test set	Training set	Test set
R1^2	0.013	0.030	0.004	0.156	0.131	0.294	0.006	0.102
R2^2	0.192	0.062	0.321	0.069	0.066	0.054	0.029	0.002
R3^2	0.233	0.376	0.192	0.280	0.056	0.182	0.225	0.018
R4^2	0.033	0.045	0.348	0.231	0.030	0.152	0.003	0.255
R5^2	0.011	0.125	0.010	0.003	0.065	0.195	0.001	0.283
R6^2	0.056	0.018	0.041	0.034	0.234	0.171	0.095	0.019
R7^2	0.364	0.113	0.213	0.465	0.013	0.14	0.003	0.273
R8^2	0.325	0.213	0.192	0.242	0.270	0.248	0.019	0.127
R9^2	0.337	0.005	0.215	0.205	0.032	0.142	0.055	0.026
R10^2	0.004	0.233	0.029	0.013	0.005	0.064	0.041	0.122
Average Rx^2	0.157	0.122	0.157	0.170	0.090	0.164	0.048	0.123
Original R^2	0.889	0.818	0.982	0.867	0.961	0.828	0.999	0.991

---

Moreover, the residuals between the predicted and experimental values for all the developed models were calculated and shown in Fig. 4 and 5. As most of the residuals are randomly distributed on both sides of the zero baseline without obvious regularity, one may conclude that there are no systematic errors exist in developing the presented models.

Fig. 4 Plots of the residuals for predicting toxicity to E. coli.

Fig. 5 Plots of the residuals for predicting toxicity to HaCaT cells.

All the results discussed above showed that the presented models positively pass internal and external validations with satisfactory stability and predictivity, which can be reliably used to predict the toxicity of MeOx NPs. Considering the limited experimental data available and the complex nature of cytotoxicity of NPs, it was not possible to further improve the model predictions beyond the current results.

This work also demonstrated that the newly proposed improved SMILES-based optimal descriptors can be successfully employed to model the complex toxicity of MeOx NPs to both prokaryotic and eukaryotic cells. Once properly developed, the nano-QSAR models can provide predictions of the toxicity of MeOx NPs quickly using only SMILES structures and some basic physicochemical properties. The SMILES derived solely from the molecular structures can be generated by freely accessible software such as ChemSketch and CORAL, while the employed physicochemical properties can be easily obtained from molecular formula and periodic table.

3.4 Comparison with previous works

Comparison of the presented models has been performed with other previously reported studies on the same endpoints. Considering the statistical parameters as demonstrated in Table 3, there is no doubt that the presented models obtained from our improved SMILES-based optimal descriptors are better than or at least comparable to those of the previous models, especially for predicting the toxicity of NPs to E. coli. Moreover, the new proposed descriptors are less computationally demanding and more reproducible than the ones used in the previously reported models. More other important characteristics of the models are further discussed below.

Comparisons between model 1 and Puzyn et al.'s²⁶ model (both models were developed based on the same dataset without considering the size effect on cytotoxicity to E. coli). Firstly, Puzyn et al.'s model employed the more interpretable mechanistically quantum-mechanical descriptors, but model 1 employing the improved SMILES-based optimal descriptors showed better predictive ability. Regarding the applicability efficiency and applicability range, model 1 need no complex quantum-mechanical calculations and thus is considered to be easier to apply. Moreover, the use of model 1 requires only knowledge of SMILES notations and some physicochemical features directly obtained from molecular formula, but the quantum-chemical descriptors employed in Puzyn et al.'s model were always calculated using simplified clusters based on certain prior assumptions, thus Puzyn et al.'s model are considered to be reliable only for the MeOx NPs with the size fall between certain ranges (usually be defined between 15–90 nm).

Comparisons between model 2 and Sizochenko et al.'s²⁸ model (both models were developed based on the same dataset considering the size effect on cytotoxicity to E. coli). Firstly, Sizochenko et al. used a set of 7 interpretable mechanistically descriptors in their model, but the resulting QSAR model appeared to suffer from the risk of overfitting. The model was obtained only from 13 training set NPs but with 7 descriptors, where the ratio of number of chemicals in the training set to that of employed descriptors is relatively low (a minimum ratio of 5 : 1 is suggested⁴⁶). If there are too few samples or too many descriptors inputting in the QSAR model, the risk of overfitting is fairly high. Regarding the model performance, our model 2 showed a better calibration ability for the training set (0.9824 vs. 0.93), as well as a much better predictive ability for the test set (0.8670 vs. 0.78). Regarding the modeling methods employed, Sizochenko et al.'s model was established by using Random Forest (RF) regression, while model 2 was developed by employing the multiple regressions (MLR) technique. Compared with the ensembled RF model which can not be visually presented, model 2 was more simple and intuitive. Considering the fact that the descriptors employed in both models are easily available, model 2 is considered to be easier to apply due to its transparent feature.

Comparisons between model 3 and Gajewicz et al.'s³⁰ model (both models were developed based on the same dataset without considering the size effect on cytotoxicity to HaCaT cells). Gajewicz et al.'s model employed two interpretable quantum-mechanical descriptors, but model 3 employing the improved SMILES-based optimal descriptors showed better predictive ability. Moreover, model 3 is conceptually simple and easy to apply, and both models are considered to theoretically be applicable to any MeOx for which the molecular structure is known.

Comparisons between model 4 and Sizochenko et al.'s²⁸ model (both models were developed based on the same dataset considering the size effect on cytotoxicity to HaCaT cells). Firstly, Sizochenko et al. used a set of 6 interpretable mechanistically descriptors for their model, but the resulting QSAR model also appeared to suffer from the risk of overfitting⁴⁹ (obtained from 14 training set NPs but with 6 descriptors). Also, model 4 showed both better calibration and predictive ability for the training (0.9997 vs. 0.96) and test (0.9905 vs. 0.92) sets. Moreover, Sizochenko et al. employed the RF method to develop a reliable but non-intuitive model, while model 4 was simpler and more transparent and can be considered to be easier to apply.

Consequently, the presented nano-QSAR models can be considered to be able to predict the toxicity of MeOx NPs to both prokaryotic and eukaryotic systems with some obvious superiority in comparison to the previously developed models.

3.5 Mechanistic interpretation

The obtained nano-QSAR models could also provide some important insight into mechanism of action of cytotoxicity to both E. coli and HaCaT cells.

There are two main types of mechanisms of MeOx NPs' toxicity reported in the literature. The first type (Mechanism I) involves the detachment of metal cations from the surface of MeOx, while the second one (Mechanism II) is related to the redox properties of the MeOx' surfaces (transfer of electrons between the surface of MeOx and intracellular redox couples).^51–53 Both processes lead to the formation of highly reactive and less specific hydroxyl radicals, which could be the main factors in charge of inducing oxidative stress in the cells.

In this study, a total of five basic physicochemical features were included into the SMILES-based optimal descriptors to improve the characterization of nano-structure information and build up predictive models. Sensitivity analysis was performed to determine the relative importance of each feature towards the toxicity of the NPs to both cells. The R² value for the new reduced model on the training set was computed when the ith feature is excluded from the original model. The process is repeated for each feature. Then differences between the R² of the original model and the reduced ones were calculated and shown as R²_diffi (see Table 5). It is obvious that the most important feature is the one that leads to the highest value of R²_diffi. The values of R²_diffi for both models were plotted in Fig. 6.

Table 5 Decreases in R² for each feature in both models

Features	EC50			LC50
	Original R^2	Reduced R^2	R^2diffi	Original R^2	Reduced R^2	R^2diffi
Molecular weight	0.9824	0.8429	0.1395	0.9997	0.8991	0.1006
Cationic charge		0.8329	0.1495		0.6018	0.3979
Mass percentage of metal elements		0.8401	0.1423		0.5657	0.4340
Individual size		0.8349	0.1475		0.5644	0.4353
Aggregation size		0.8130	0.1694		0.5428	0.4569

---

Fig. 6 Sensitivity analysis results of both nano-QSAR models.

According to the results of sensitivity analysis, the features are ranked according to their corresponding relative importance as follows:

(1) For QSAR model of toxicity to E. coli: aggregation size > cationic charge > individual size > mass percentage of metal elements > molecular weight;

(2) For QSAR model of toxicity to HaCaT cells: aggregation size > individual size > mass percentage of metal elements > cationic charge > molecular weight.

The individual size and aggregation size were found to be the most important factors for the cytotoxicity to both E. coli and HaCaT cell, especially for the latter. It can thus be concluded that the size effects of MeOx NPs on their cytotoxicity were still found to be significant for NPs with sizes studied here (between 15 and 150 nm), at least for the HaCaT cells. That is to say, the particle size did affect activity for the NPs with sizes between 15 and 150 nm. However, most previous studies revealed that the most significant size-dependent changes of toxicity of the NPs occur below about 5 nm, while for NPs with sizes between 15 and 90 nm, the toxicity changes are always neglected. The possible reasons for the discrepancy may be attributed to the fact that the literature findings may be significant for the E. coli (prokaryotic cells), while for the eukaryotic HaCaT cells, it may be not obvious. Also, as has been stated in previous studies, the toxicity of MeOx increases with a decreasing size.^54–56 It can be explained by both mechanisms discussed above. Firstly, along with the increase of the ratio of surface area to volume, the total number of atoms at the NP surface will increase. Moreover, because the surface atoms will form fewer chemical bonds than the interior ones, they are less stable and could be more easily detached from the surface. This process is in agreement with Mechanism I of nanotoxicity discussed above. Meanwhile, as very small NPs behave more like single molecules instead of bulk crystals, the differences in physicochemical properties (including redox properties) between the NPs and the bulk should be significant, especially for those of very small sizes, for which large fractions of atoms are present at the surface. This process is in agreement with Mechanism II discussed above. However, it should also be pointed out that, the toxicity of MeOx increases with a decreasing size, but become insignificant at sizes above certain values (dependent on the particular MeOx).

The cationic charge is another important physicochemical feature to interpret and predict the cytotoxicity of MeOx. As can be seen from Table S1,† those oxides with low cationic charge (like CuO, ZnO, CoO and NiO) usually indicate strong reductive properties (easy detachment of the metal cation), which enhance their cytotoxicity to the E. coli. That is to say, with an increase in the cationic charge of MeOx NPs, the cytotoxicity to E. coli will decrease for most MeOx NPs. This mechanism is in agreement with Mechanism I of toxicity discussed above. As for the HaCaT cells, there is no obvious relationship between the cytotoxicity and cationic charge for most MeOx NPs, which indicated the different mechanism of toxicity existed between prokaryotic and eukaryotic systems.

The relative contributions of mass percentage of metal elements are different in both models. Compared to E. coli, this feature, which is related to the number of metal atoms in a crystal, has a relatively higher impact on HaCaT cells. The cytotoxicity to both cells will decrease with an increase in the mass percentage of metal elements for MeOx NPs.

The comparison studies of the toxicity of MeOx NPs to both cells verified the hypothesis that different modes of toxic action occur between prokaryotic and eukaryotic systems.

4 Conclusions

In this study, a novel kind of nanospecific theoretical descriptor has been proposed for nano-QSAR modeling on the cytotoxicity of MeOx NPs to both E. coli and HaCaT cells. Corresponding robust and predictive nano-QSAR models have been successfully developed and rigorously validated. Also, the predominant nano-structure features responsible for cytotoxicity of MeOx NPs were identified and differences in the mechanisms of their toxicity to different types of cells were discussed. The major findings of this particular study are summarized below.

(1) The new developed descriptors can be obtained directly from SMILES and some common physicochemical features without computationally demanding calculations requiring any quantum-chemical background. These reproducible descriptors can also efficiently encode cytotoxicity of MeOx NPs leading to models with high statistical quality as well as certain interpretability. It could be assumed that the new descriptors could be further applied to other unexplored inorganic compounds.

(2) We have successfully developed a set of statistically significant and robust nano-QSAR models for predicting toxicity of MeOx NPs to both the E. coli bacteria and human keratinocytes. The presented models were also conceptually simple, computationally inexpensive, easy to apply, and with acceptable model interpretability, when comparing to the literature models.

(3) The individual size and aggregation size were found to be the most important factors for the cytotoxicity to both E. coli and HaCaT cells, especially for the latter. Also, MeOx having low cationic charge usually exhibit the strong reductive properties, which enhances their cytotoxicity to the E. coli bacteria and vice versa. The developed nano-QSAR models revealed and verified the differences in the mechanisms of toxicity of MeOx NPs to prokaryotic and eukaryotic systems.

(4)This study can provide a new way for predicting cytotoxicity for new MeOx NPs or for other NPs for which experimental values are unknown. Also, the developed models have potential applications in improving the experimental design of safer NPs and enabling their prioritization for virtual screening.

Acknowledgements

This research was supported by National Natural Science Fund of China (no. 21436006, 21576136), and Natural Science Fund of Jiangsu Higher Education Institutions of China (no. 12KJA620001). The authors would like to thank Prof. Andrey A. Toropov for all of his suggestion on the development of nano-structure descriptors.

Notes and references

1. M. Ouyang, J. L. Huang and C. M. Lieber, Acc. Chem. Res., 2002, 35, 1018–1025.
2. T. Puzyn, D. Leszczynska and J. Leszczynski, Small, 2009, 5, 2494–2509.
3. A. Nel, T. Xia, L. Mädler and N. Li, Science, 2006, 311, 622–627.
4. M. Auffan, J. Rose, J. Y. Bottero, G. V. Lowry, J. P. Jolivet and M. R. Wiesner, Nat. Nanotechnol., 2009, 4, 634–641.
5. R. Jones, Nat. Nanotechnol., 2009, 4, 75.
6. M. D. Newman, M. Stotland and J. I. Ellis, J. Am. Acad. Dermatol., 2009, 61, 685–692.
7. J. M. Nam, C. C. Thaxton and C. A. Mirkin, Science, 2003, 301, 1884–1886.
8. R. Mahtab, J. P. Rogers and C. J. Murphy, J. Am. Chem. Soc., 1995, 117, 9099–9100.
9. I. H. El-Sayed, X. Huang and M. A. El-Sayed, Cancer Lett., 2006, 239, 129–135.
10. M. Lewin, N. Carlesso, C. H. Tung, X. W. Tang, D. Cory, D. T. Scadden and R. Weissleder, Nat. Biotechnol., 2000, 18, 410–414.
11. N. L. Rosi, D. A. Giljohann, C. S. Thaxton, A. K. Lytton-Jean, M. S. Han and C. A. Mirkin, Science, 2006, 312, 1027–1030.
12. G. Han, C. C. You, B. J. Kim, R. S. Turingan, N. S. Forbes, C. T. Martin and V. M. Rotello, Angew. Chem., 2006, 118, 3237–3241.
13. J. Dong and J. I. Zink, ACS Nano, 2014, 8, 5199–5207.
14. A. A. Hwang, J. Lu, F. Tamanoi and J. I. Zink, Small, 2015, 11, 319–328.
15. H. Meng, T. Xia, S. George and A. E. Nel, ACS Nano, 2009, 3, 1620–1627.
16. Y. Bai, Y. Zhang, J. Zhang, Q. Mu, W. Zhang, E. R. Butch and B. Yan, Nat. Nanotechnol., 2010, 5, 683–689.
17. A. D. Maynard, R. J. Aitken, T. Butz, V. Colvin, K. Donaldson, G. Oberdörster, M. A. Philbert, J. Ryan, A. Seaton, V. Stone, S. S. Tinkle, L. Tran, N. J. Walker and D. B. Warheit, Nature, 2006, 444, 267–269.
18. R. F. Service, Science, 2004, 304, 1732.
19. A. R. Katritzky, V. S. Lobanov and M. Karelson, Chem. Soc. Rev., 1995, 24, 279–287.
20. J. Taskinen and J. Yliruusi, Adv. Drug Delivery Rev., 2003, 55, 1163–1183.
21. Y. Pan, J. C. Jiang, X. Y. Ding, R. Wang and J. J. Jiang, AIChE J., 2010, 56, 690–701.
22. A. R. Katritzky, M. Kuanar, S. Slavov, C. D. Hall, M. Karelson, I. Kahn and D. A. Dobchev, Chem. Rev., 2010, 110, 5714–5789.
23. P. R. Gil, G. Oberdörster, A. Elder, V. Puntes and W. J. Parak, ACS Nano, 2010, 4, 5527–5531.
24. D. A. Winkler, E. Mombelli, A. Pietroiusti, L. Tran, A. Worth, B. Fadeel and M. J. McCall, Toxicology, 2013, 313, 15–23.
25. T. Le, V. C. Epa, F. R. Burden and D. A. Winkler, Chem. Rev., 2012, 112, 2889–2919.
26. T. Puzyn, B. Rasulev, A. Gajewicz, X. Hu, T. P. Dasari, A. Michalkova, H. Hwang, A. Toropov, D. Leszczynska and J. Leszczynski, Nat. Nanotechnol., 2011, 6, 175–178.
27. A. A. Toropov, A. P. Toropova, E. Benfenati, G. Gini, T. Puzyn, D. Leszczynska and J. Leszczynski, Chemosphere, 2012, 89, 1098–1102.
28. N. Sizochenko, B. Rasulev, A. Gajewicz, V. Kuz'min, T. Puzyn and J. Leszczynski, Nanoscale, 2014, 6, 13986–13993.
29. S. Kar, A. Gajewicz, T. Puzyn, K. Roy and J. Leszczynski, Ecotoxicol. Environ. Saf., 2014, 107, 162–169.
30. A. Gajewicz, N. Schaeublin, B. Rasulev, S. Hussain, D. Leszczynska, T. Puzyn and J. Leszczynski, Nanotoxicology, 2014, 8, 1–13.
31. S. Kar, A. Gajewicz, T. Puzyn and K. Roy, Toxicol. in Vitro, 2014, 28, 600–606.
32. M. Ghorbanzadeh, M. H. Fatemi and M. Karimpour, Ind. Eng. Chem. Res., 2012, 51, 10712–10718.
33. D. Fourches, D. Pu, C. Tassa, R. Weissleder, S. Y. Shaw, R. J. Mumper and A. Tropsha, ACS Nano, 2010, 4, 5703–5712.
34. A. A. Toropov, A. P. Toropova, T. Puzyn, E. Benfenati, G. Gini, D. Leszczynska and J. Leszczynski, Chemosphere, 2013, 92, 31–37.
35. V. C. Epa, F. R. Burden, C. Tassa, R. Weissleder, S. Shaw and D. A. Winkler, Nano Lett., 2012, 12, 5808–5812.
36. A. P. Toropova and A. A. Toropov, Chemosphere, 2013, 93, 2650–2655.
37. A. A. Toropov, B. F. Rasulev, D. Leszczynska and J. Leszczynski, Chem. Phys. Lett., 2008, 457, 332–336.
38. H. S. Choi, Y. Ashitate, J. H. Lee, S. H. Kim, A. Matsui, N. Insin, M. G. Bawendi, M. Semmler-Behnke, J. V. Frangioni and A. Tsuda, Nat. Biotechnol., 2010, 28, 1300–1303.
39. Y. Pan, S. Neuss, A. Leifert, M. Fischler, F. Wen, U. Simon, G. Schmid, W. Brandau and W. Jahnen-Dechent, Small, 2007, 3, 1941–1949.
40. L. Kukreja, S. Barik and P. Misra, J. Cryst. Growth, 2004, 268, 531–535.
41. Z. W. Qu and G. J. Kroes, J. Phys. Chem. B, 2006, 110, 8998–9007.
42. H. J. Zhai and L. S. Wang, J. Am. Chem. Soc., 2007, 129, 3022–3026.
43. L. K. Adams, D. Y. Lyon and P. J. Alvarez, Water Res., 2006, 40, 3527–3532.
44. N. Chirico and P. Gramatica, J. Chem. Inf. Model., 2011, 51, 2320–2335.
45. P. Gramatica, QSAR Comb. Sci., 2007, 26, 694–701.
46. A. Tropsha, P. Gramatica and V. K. Gombar, QSAR Comb. Sci., 2003, 22, 69–77.
47. R. D. Clark, D. G. Sprous and J. M. Leonard, Validating models based on large dataset, Proceedings of the 13th European Symposium on Quantitative Structure–Activity Relationships, 2001.
48. S. Wold and L. Eriksson, in Chemometric Methods in Molecular Design, ed. H. van de Waterbeend, VCH, New York, 1995, pp. 309–318.
49. S. Kovarich, E. Papa and P. Gramatica, J. Hazard. Mater., 2011, 190, 106–112.
50. A. Golbraikh and A. Tropsha, J. Mol. Graphics Modell., 2002, 20, 269–276.
51. E. Burello and A. P. Worth, Nanotoxicology, 2011, 5, 228–235.
52. E. Burello and A. P. Worth, Wiley Interdiscip. Rev.: Nanomed. Nanobiotechnol., 2011, 3, 298–306.
53. H. Zhang, Z. Ji, T. Xia, H. Meng, C. Low-Kam, R. Liu, S. Pokhrel, S. Lin, X. Wang, Y. P. Liao, M. Wang, L. Li, R. Rallo, R. Damoiseaux, D. Telesca, L. Mädler, Y. Cohen, J. I. Zink and A. E. Nel, ACS Nano, 2012, 6, 4349–4368.
54. C. Carlson, S. M. Hussain, A. M. Schrand, L. K. Braydich-Stolle, K. L. Hess, R. L. Jones and J. J. Schlager, J. Phys. Chem. B, 2008, 112, 13608–13619.
55. M. Kumari, S. Rajak, S. P. Singh, S. I. Kumari, P. U. Kumar, U. S. Murty, M. Mahboob, P. Grover and M. F. Rahman, J. Nanosci. Nanotechnol., 2012, 12, 2149–2159.
56. B. M. Prabhu, S. F. Ali, R. C. Murdock, S. M. Hussain and M. Srivatsan, Nanotoxicology, 2010, 4, 150–160.

---

Footnote

† Electronic supplementary information (ESI) available: The experimental toxicity values as well as corresponding physicochemical feature values for each metal oxide nanoparticle. See DOI: 10.1039/c6ra01298a

---