Evaluating the cytotoxicity of a large pool of metal oxide nanoparticles to Escherichia coli: Mechanistic understanding through In Vitro and In Silico studies

Abstract

The toxic effect of eight metal oxide nanoparticles (MONPs) on Escherichia coli was experimentally evaluated following standard bioassay protocols. The obtained cytotoxicity ranking of these studied MONPs is Er₂O₃, Gd₂O₃, CeO₂, Co₂O₃, Mn₂O₃, Co₃O₄, Fe₃O₄/WO₃ (in descending order). The computed EC₅₀ values from experimental data suggested that Er₂O₃ and Gd₂O₃ were the most acutely toxic MONPs to E. coli. To identify the mechanism of toxicity of these 8 MONPs along with 17 other MONPs from our previous study, we employed seven classifications and machine learning (ML) algorithms including linear discriminant analysis (LDA), naïve bayes (NB), multinomial logistic regression (MLogitR), sequential minimal optimization (SMO), AdaBoost, J48, and random forest (RF). We also employed 1^st and 2^nd generation periodic table descriptors developed by us (without any sophisticated computing facilities) along with experimentally analyzed Zeta-potential, to model the cytotoxicity of these MONPs. Based on qualitative validation metrics, the LDA model appeared to be the best among the 7 tested models. The core environment of metal defined by the ratio of the number of core electrons to the number of valence electrons and the electronegativity count of oxygen showed a positive impact on toxicity. The identified properties were important for understanding the mechanisms of nanotoxicity and for predicting the potential environmental risk associated with MONPs exposure. The developed models can be utilized for environmental risk assessment of any untested MONP to E. coli, thereby providing a scientific basis for the design and preparation of safe nanomaterials.

Graphical Abstract

1. Introduction

An array of metal oxide nanoparticle products exists in the market due to the globalization of various segments including biomedical and healthcare products, car productions, aerospace, and aviation industries, paint manufacturing, fine to bulk chemical industries, batteries, as well as their applications in information and communication technologies (Gajewicz et al., 2012; Nowack et al., 2013; Pathakoti et al., 2018, 2019). According to a recent report from Research and Markets, the possible global nanoparticle drug market may reach over US$ 200 Billion by 2024, which represents approximately 10% of growth as per compound annual growth rate (CAGR) (Global Nanoparticle Drug Delivery Market, Dosage, Price and Clinical Pipeline Outlook, 2024, 2020). Thus, along with the expanded global market, there is a constant threat of nanoparticle-induced toxicity to human health and the ecosystem. Among nanoparticle-based products, around 80% consists of metal oxides in which 95% are ZnO, TiO₂, Fe₂O₃, Al₂O_3, and SiO₂. This ranks them among the most vulnerable compounds. Thus, a complete toxicity profile in form of cell viability and cytotoxicity (apoptosis, membrane damage, cell proliferation), oxidative stress (ROS production, lipid peroxidation), pro-inflammatory response (cytokine production), genotoxicity (DNA-damage) needs to be analyzed for MONPs risk assessment and safety evaluation (Forest et al., 2019).

As it is impossible to examine thousands of MONP-based commercial products through in vivo and in vitro studies, in silico approaches are one of the best possible alternatives to save time, money, and most importantly animal sacrifice (Fourches et al., 2010). The in silico approaches are widely accepted by regulatory agencies over the world, including the United States Environmental Protection Agency (U.S. EPA) for the toxicity data gap filling of those MONPs which have no experimental data available. Among in silico approaches, quantitative structure-activity relationship (QSAR) (Kar et al., 2014a, 2019; Ojha et al., 2019; Puzyn et al., 2011; Winkler et al., 2013), machine learning (ML) (Gajewicz et al., 2018; Fjodorova et al., 2017), and read-across (Gajewicz, 2017; Gajewicz et al., 2015; Lamon et al., 2019) tools are most commonly employed by researchers to model the diverse toxicity endpoints. The toxic effects of MONPs have been tested on prokaryotic cells (Aruoja et al., 2015; Gajewicz, 2017; Kar et al., 2014b; Pathakoti et al., 2014; Puzyn et al., 2011; Sizochenko et al., 2016; Speck-Planche et al., 2015; Toropova et al., 2016), eukaryotic cells (Basant and Gupta, 2017a; Epa et al., 2012; Esposito et al., 2015; Fourches et al., 2010, 2016; Ivask et al., 2015; Kar et al., 2014; Lin et al., 2013; Liu et al., 2015; Ojha et al., 2019; Papa et al., 2016; Shin et al., 2017; Winkler et al., 2013, 2014; Zhang et al., 2012) or both (Basant and Gupta, 2017b; Concu et al., 2017; De et al., 2018; Kleandrova et al., 2014).

Puzyn et al. (2011) developed a multiple linear regression (MLR) model for the cytotoxicity (in EC₅₀ which represents the effective concentration of a given oxide that reduces bacterial viability by 50%) towards Escherichia coli (E. coli) employing 17 MONPs with the use of all quantum-chemical descriptors and found that only one property called ΔH_Me+, the enthalpy of formation of a gaseous cation, can predict 86% of correlation with the cytotoxicity. Kar et al. (2014b) used periodic table based descriptors to build QSAR model using the same dataset and hypothesized that MONPs release an electron easily compared to the crystal structure of the same particles and initiate the formation of reactive oxygen species, which constitute the root cause of oxidative stress condition to E. coli and subsequent toxicity. Mu et al. (2016) reported that the enthalpy of formation of a gaseous cation (ΔH_Me+) and polarization force (Z/r) is the best possible feature to model the same dataset. Basant and Gupta (2017b) illustrated enthalpy of formation of a gaseous cation and oxygen in weight percentage to model cytotoxicity employing a random forest classification model. Singh and Gupta (2014) prepared a highly predictive ensemble learning model using molar refractivity, polar surface area, and oxygen percent. Sizochenko et al. (2014) modeled the same dataset employing liquid drop model (LDM)-based descriptors, while Fjodorova et al. (2017) employed artificial neural network models to correlate the cytotoxicity of MONPs using the number of metal atoms in oxide, the number of oxygen atoms in oxide, charge of a metal cation and metal electronegativity by Pauling scale. Zhou et al. (2017) modeled cytotoxicity of MONPs to E. coli with the DFT derived (B3LYP method) quantum-chemical descriptors using the MLR and support vector machine (SVM) methods to reveal the importance of the lowest unoccupied molecular orbital (LUMO) energy and molar heat capacity (Cp) properties for the development of a predictive model.

It is interesting to point out that more than 20 QSAR models were developed employing the same dataset (17 MONPs) generated by Puzyn et al. (2011) using multiple statistical methods and in most of the cases employing quantum-chemical descriptors. If we summarize the interpretation of cytotoxicity of MONPs towards E. coli considering 17 MONPs, then the major identified mechanistic interpretations include the following: a) Enthalpy of formation of a gaseous cation, b) metal electronegativity and charge of the metal cation corresponding to a given oxide, c) HOMO-LUMO energy gap, and d) electrophilicity index. Kaweeteerawat et al. (2015) modeled cytotoxicity with a different measurement of half-maximal growth inhibitory concentration (IC₅₀) with a different dataset of 24 MONPs employing the classification-based support SVM and found that the conduction band energy and hydration enthalpy were two significant features to predict the toxicity endpoint to E. coli.

The present experimental dataset is one of the biggest available MONPs’ in vitro cytotoxicity (EC₅₀: effective concentration that reduces bacterial viability by 50%) data for E. coli at present time, performed under similar experimental conditions and in the same laboratory following a defined algorithm as performed by previous work (Puzyn et al., 2011). In most of the previous studies, regression- (Kar et al., 2014; Mu et al., 2016; Pan et al., 2016; Puzyn et al., 2011) and classification-based (Fjodorova et al., 2017; Sizochenko et al., 2014; Zhou et al., 2017) quantitative structure-activity relationship (QSAR) models were generated for the development of in silico models and they were quite successful. As the present dataset comprises diverse forms of MONPs, it provides an opportunity to check different methods to develop in silico models. Hence, we have considered classification based QSAR method and diverse machine learning (ML) approaches in the present study. All the models are prepared following the OECD principles and stringent validation techniques. The significance and major aims of the present study are as follows:

• In vitro cytotoxicity (EC₅₀) evaluation of eight MONPs (Er₂O₃, Gd₂O₃, CeO₂, Co₂O₃, Mn₂O₃, Co₃O₄, Fe₃O_4, and WO₃) for 2 h of exposure under dark conditions.

• Generating a cytotoxicity database of 25 MONPs by merging previous toxicity data of 17 MONPs with those of 8 new MONPs following OECD principle 1 of ‘defined endpoint’; as test algorithms are the same, and tests are performed under the same laboratory conditions.

• Developing a predictive classification based QSAR model as well as ML models for prediction of untested along with newly generated MONPs in the future.

• Previously developed in silico models were prepared and mechanistic interpretation for the cytotoxicity was generated employing 17 MONPs. Thus, one important aspect of the current research as to determine whether mechanistic interpretations for toxicity are still the same or to assess whether the introduction of an additional set of 8 MONPs (including 3 lanthanides) in the modeling offers a new mechanism of action for toxicity to E. coli.

• Taking advantage of the fact that the reliability of the statistically validated model is better than the previously developed ones due to the presence of more data points in the present model.

2. Materials and methods

2.1. Nanoparticles for study

Eight MONPs including 3 lanthanides were considered for in vitro cytotoxicity (EC₅₀) evaluation to E. coli. The studied MONPs included Er₂O₃, Gd₂O₃, CeO₂, Co₂O₃, Mn₂O₃, Co₃O₄, Fe₃O_4, and WO₃. Seventeen MONPs were taken from our previous study (Puzyn et al., 2011) whose cytotoxicity was checked following completely similar experimental protocols within the same laboratory to produce a robust dataset of 25 MONPs. The reason for the choice of these 8 MONPs was based on their widespread usage in numerous applications in day to day life followed by their potential release into the environment (Keller et al., 2013; Arvidsson et al., 2018). Among the lanthanides, CeO₂ MONPs are widely used for energy storage applications. They also have a potential antimicrobial activity (Blinova et al., 2020). Nearly 3% of the produced CeO₂ MONPs are released into the aquatic environments and largely accumulate in the sediments (Keller and Lazareva, 2013). Likewise, substantial gadolinium concentrations are reported in the aquatic environments (Hatie et al., 2016). More specifically, the increasing concentrations of MONPs in the environment may cause adverse effects to human health and environment (Cassee et al., 2011; Blinova et al., 2020). Thus, all the studied MONPs are potential contaminants to the ecosystem and need to be carefully evaluated for their environmental safety.

2.2. Physicochemical characterization of the MONPs

The primary particle size of MONPs was measured by transmission electron microscopy (TEM) (Jeol, JEM 1011 electron microscope, equipped with Gatan camera model 785). The samples were prepared by drop-coating the 5 μL of the MONP suspension on carbon-coated copper TEM grids (Ted Pella, CA) and drying overnight at room temperature. The zeta potential and particle size distribution of the MONPs (100 μg/mL, pH 7.5) in Milli-Q water were measured using a Zetasizer (Malvern Instruments).

2.3. Cell viability assay

Cell viability experiments were determined with E. coli (ATCC#25254) using the colony forming units (CFU) method on Luria-bertani (LB) agar petri plates. The exposure duration was 2 h and both the control and treated samples were agitated at 150 rpm. Preliminary tests were carried out with a wide range of concentrations initially (1–1000 ppm) and followed by the definitive toxicity tests that were conducted using four to five concentrations causing 10e90% decrease in cell viability. All tests were completed within six months. The cytotoxicity of the tested MONPs was expressed in terms of EC₅₀ (the MONP concentration that causes 50% mortality of the tested bacterium, E. coli). The EC₅₀ values were calculated from experimental data using linear regression in Excel.

2.4. Dataset for in silico modeling

The prepared dataset consisted of 25 MONPs with cytotoxicity data in form of EC₅₀ expressed in ppm. For modeling purposes, all toxicity values were transformed into a negative logarithmic scale expressed in the molar scale. As mentioned earlier, all toxicity values were checked through similar protocols and under the same laboratory conditions. Thus, the dataset complied with the OECD principle 1 for in silico modeling. Out of 25 MONPs, two including Fe₃O₄ and WO₃ did not show any toxicity even at the highest tested concentration of 2000 ppm. Therefore, their toxicity could not be quantified. On the contrary, Gd₂O₃ and Er₂O₃ exhibited quite a higher toxicity compared to the remaining MONPs.

2.5. Descriptors for modeling

Seven 1^st generation (Kar et al., 2014b) and sixteen 2^nd generation (De et al., 2018) periodic table-based descriptors were computed. Experimentally generated Zeta Potential (mV) was also considered as a descriptor as it is one of the proven descriptors for modeling study of MONPs. Thus, a total pool of 24 descriptors was employed for the modeling purpose. Periodic table descriptors showed potential impact for modeling of MONPs over the years and most importantly these can be calculated without any software as well as save a huge amount of computation time, compared to approaches involving other quantum descriptors. The complete list of the employed descriptors is reported in Table 1.

Table 1.

List of descriptors used for model development.

No.	Type of descriptors	Mathematical Expression	Description
1	1st Generation periodic table-based	MW	Molecular weight of metal oxide
2	descriptors	Nmetal	Number of metal
3		Noxy	Number of oxygen
4		χ	Metal electronegativity
5		$\sum \chi$	Total metal electronegativity in specific metal oxide
6		$\sum \chi /\mathrm{nO}$	Total metal electronegativity in specific metal oxide relative to number of oxygen
7		χox	Oxidation number of metal
8	2nd Generation periodic table-based	Zmetal	Atomic number of metal
9	descriptors	Zvmetal	Valence electron of metal
10		PNmetal	Period number of metal
11		λ = (Zmetal − Zvmetal)/Zvmetal	Core environment of metal defined by the ratio of the number of core electrons to the number of valence electrons
12		μ = 1/(PNmetal)	–
13		Vmetal	Valence of metal
14		αmetal = λ*μ	–
15		$\sum {\alpha }_{\mathrm{metal}}={\alpha }_{\mathrm{metal}}*{\mathrm{N}}_{\mathrm{metal}}$	–
16		$\sum {\alpha }_{\mathrm{oxy}}={\mathrm{N}}_{\mathrm{oxy}}*\text{0.33}$	–
17		$\sum \alpha =\sum {\alpha }_{\mathrm{metal}}+\sum {\alpha }_{\mathrm{oxy}}$	The core count, gives a measure of the molecular bulk
18		εmetal = −αmetal +(0.3*Zvmetal)	Electronegativity count of metal
19		εoxy = −αoxy + (0.3*Zvoxy)	Electronegativity count of oxygen
20		$\sum \mathrm{\epsilon }={\mathrm{\epsilon }}_{\mathrm{metal}}*{\mathrm{N}}_{\mathrm{metal}}+{\mathrm{\epsilon }}_{\mathrm{oxy}}*{\mathrm{N}}_{\mathrm{oxy}}$	Total electronegativity count of MONP
21		$\sum \mathrm{\epsilon }/\mathrm{N}$	Sum epsilon relative to number of atoms in the molecule
22		$(\sum \mathrm{\alpha }{)}^2$	Square of summation of alpha, gives measure of molecular bulk
23		$(\sum \mathrm{\epsilon }/\mathrm{N}{)}^2$	Squared sum epsilon by number of atoms
24	Experimental descriptors	ζ-potential	Zeta-potential of MONP

2.6. Dataset division and in silico modeling tools

The dataset was divided into training/calibration and test/validation sets employing Euclidean distance-based method employing Dataset Division GUI 2.1 software (http://teqip.jdvu.ac.in/QSAR_Tools/). The training and test sets consisted of 17 and 8 MONPs, respectively where training set was used for model development and calibration, and test set was employed for validation of the developed model.

To develop a classification based QSAR model, we employed linear discriminant analysis (LDA) (Xanthopoulos et al., 2013) using the in house script of R-Studio (RStudio Team (2020). Also, six classification-based ML approaches were used to have a total of seven models to compare the statistical quality as well as to identify the best model for prediction purposes and to understand and interpret the major mechanism of toxicity. These six ML approaches included naïve bayes (NB) (Cheeseman and Stutz, 1996), sequential minimal optimization (SMO) (Osuna et al., 1997), AdaBoost (Bartlett and Traskin, 2007), multinomial logistic regression (MLogitR) (Greene, 2012), random forest (RF) (Ho, 1998) and J48 (Quinlan, 1993) which were implemented under the WEKA workbench (Frank et al., 2016). Based on the normality distribution plot for log(1/EC₅₀) toxicity values, MONPs with ≤3.14 toxicity values were considered highly toxic and denoted as H, while the remaining ones were considered to have lower toxicity (denoted as L). As WO₃ and Fe₃O₄ elucidated no toxicity up to 2,000 ppm, they were also put in the lower toxicity group.

2.7. Model validation

The classification-based models were validated using classical qualitative metrics such as Sensitivity, Specificity, Accuracy, Precision, and F-measure. Based on the initial assessment of these metrics, and the subsequent comparison of all seven models, we identified the best classification model. Thereafter, the best model was checked further with stringent validation metrics based on Matthews correlation coefficient (MCC), squared Mahalanobis distance, geometric means (G-means), Cohen’s kappa, and receiver operating characteristic (ROC) curve (Roy and Kar, 2015). The model was also judged by relevant statistical tests such as Wilks’ λ statistics, chi-square (χ²), and Canonical index (R_c) (Roy and Kar, 2015). To judge the quality of the model as well as to indicate higher differentiation between two classes and the independence, a diagnostic variable was computed by calculating the area under the ROC curve (AUROC). In an ideal case, the AUROC should be 1, whereas a random case will offer a value of 0.5. The ROC graph Euclidean distance (ROCED) and the ROC graph Euclidean distance corrected with Fitness Function (FIT(λ)) (ROCFIT) were computed (Perez-Garrido et al., 2011). Finally, to identify the extent of discrimination and degree attained in the calibration or training as well as test or validation set observations, a pharmacological distribution diagram (PDD) (Murcia-Soler et al., 2003) was prepared for the visualization purpose. The best model was tested with the applicability domain (AD) test employing the Euclidean distance based method employing Euclidean Applicability domain 1.0 software (http://teqip.jdvu.ac.in/QSAR_Tools/).

3. Results and discussion

3.1. Experimental In vitro study

3.1.1. Nanoparticle characterization

All the eight MONPs were characterized and their physicochemical properties such as hydrodynamic size in H₂O (nm), Zeta potential (mV) were presented in Table 2. The average particle sizes as measured by the TEM ranged from 13 to 81 nm. As illustrated on Fig. 1, the TEM images of all the 25 MONPs, show large aggregations. Also, the primary sizes for Co₂O₃ and Gd₂O₃ were not measured by TEM due to their large agglomeration (Fig. 1). Most of the MONPs showed a positive surface charge in distilled water, whereas, Fe₃O₄ and WO₃ exhibited a negative surface charge.

Table 2.

Physicochemical characterization and EC₅₀ values of MONPs.

Metal Oxide	Particle size (vendor) (nm)	Particle size TEM (nm)	Hydrodynamic Size in H2O (nm)	Zeta Potential (mV)	Surface area (m2/g)	EC50 (ppm)
CeO2	25	17–23	197.6	12.97 ± 0.55	NA	64.05
Co2O3	<100	<100	NA	7.21 ± 1.08	>10	77.145
Er2O3	40–50	16–30	NA	9.32 ± 0.54	16	0.033
Fe3O4	10–100	23–42	198.4	−9.05 ± 0.54	NA	>2000
Gd2O3	<100	<100	195.7	20.97 ± 0.65	10–40	1.112
Mn2O3	40–60	13–20	268.8	11.03 ± 0.12	∼13.5	130
WO3	30–70 nm	26–81	176.6	−9.39 ± 0.31	NA	>2000
Co3O4	51–132	78	174.5	24.6	NA	240

^*NA: not available.

Fig. 1.

TEM images of all 25 MONPs.

3.1.2. Toxicity data

The EC₅₀ values characterizing the acute toxicity of the eight tested MONPs on E. coli after 2 h exposure under dark conditions is presented in Table 2. As shown in this figure, Er₂O₃ and Gd₂O₃ were highly toxic to E. coli with EC₅₀ values of 0.033 and 1.112 ppm, respectively. CeO₂ and Co₂O₃ were moderately toxic and Mn₂O₃ was least toxic to E. coli among the newly studied MONPs. Furthermore, WO₃ and Fe₃O₄ did not show toxicity even at 2000 ppm, the highest tested concentration in this study.

To perform the computational modeling, the toxicity of newly studied eight MONPs was merged with previously performed experimental toxicity data (Puzyn et al., 2011) of 17 MONPs tested experimentally in the same laboratory (Table 3).

Table 3.

Experimental and predicted classification toxicity data of modeled MONPs.

ID	Metal Oxides	Experimental EC50 (ppm)	Experimental pEC50 (molar)	Classification based QSAR
				Observed classification	Predicted Classification with LDA
1	Al2O3	329.936	2.49	L	L
2	CoO	23.156	3.51	H	H
3a	Cr2O3	469.691	2.51	L	H
4	CuO	50.190	3.2	H	H
5	Fe2O3	818.973	2.29	L	L
6	In2O3	430.003	2.81	L	L
7a	La2O3	439.503	2.87	L	L
8a	NiO	26.502	3.45	H	H
9	Sb2O3	667.827	2.64	L	L
10	SiO2	379.098	2.2	L	L
11a	SnO2	1472.775	2.01	L	L
12	TiO2	1453.304	1.74	L	L
13	Y2O3	304.608	2.87	L	L
14	ZnO	28.874	3.45	H	H
15	ZrO2	872.345	2.15	L	L
16a	Bi2O3	705.256	2.82	L	L
17	V2O3	108.578	3.14	L	L
18	CeO2	64.050	3.43	H	H
19a	Gd2O3	1.112	5.51	H	H
20	Mn2O3	130.000	3.08	L	L
21a	Er2O3	0.033	7.06	H	H
22	Co2O3	77.145	3.33	H	L
23a	Co3O4	240.000	3	L	L
24	WO3	>2000	NA	L	L
25	Fe3O4	>2000	NA	L	L

^aTest set compounds for LDA based classification model.

3.2. In silico modeling

3.2.1. Classification based QSAR and machine learning models

Seven classification-based in silico models are developed employing training or calibration set using 17 MONPs and validated based on the test set that comprised 8 MONPs. The qualitative results for all seven models can be found in Table 4 while the visual comparison is illustrated in Fig. 2. Considering all five metrics under validation for the training set, the top 3 developed models are MLogitR = AdaBoost > LDA. Once the developed model with the training set applied for test set validation, three out of 5 validation metrics failed miserably for the MLogitR, while AdaBoost reported 1 failed metrics along with 3 below or equal to 75% correct results. On the contrary, the LDA model offered 5 metrics with 75% or more correctness including 100% sensitivity. It is important to mention that, this is the only model which has 75% or more correctness for all five metrics for both training and test sets including 3 ideal 100% correctness and 4 reported 85% correctness. Most importantly, except for the LDA model, all remaining models had at least one failed validation metric for the test set. The idea of the best model relies on the balance of good statistical quality of validation metrics for both training and test sets which can be obtained by the LDA model which is selected for further analysis.

Table 4.

Qualitative prediction of classification based QSAR and machine learning models.

Method	Set	Sensitivity (%)	Specificity (%)	Precision (%)	Accuracy (%)	F-measure (%)
LDA	Training	80	100	100	94.1	88.9
	Test	100	80	75	87.5	85.7
NB	Training	80	91.7	80	88.2	80
	Test	100	71.4	33.3	75	50
MLogitR	Training	100	100	100	100	100
	Test	33.3	60	33.3	50	33.3
SMO	Training	100	85.7	60	88.2	75
	Test	100	71.4	33.3	75	50
AdaBoost	Training	100	100	100	100	100
	Test	100	71.4	33.3	75	50
J48	Training	100	85.7	60	88.2	75
	Test	100	71.4	33.3	75	50
RF	Training	100	85.7	60	88.2	75
	Test	100	71.4	33.3	75	50

Failed metric values are denoted in bold.

Fig. 2.

Comparison of classical validation metrics outcome among seven classification-based models.

3.2.2. Validation of the best model

The LDA analysis was performed employing a stepwise variable selection method with objective function F = 4 for inclusion; F = 3.9 for exclusion, followed by a tolerance value of 0.001 with 50% probability level. The discriminant function ΔP is represented with the following equation:

$$\begin{gathered} \Delta P=21.85\times {\epsilon }_{0xy}+19.05\times \sum {\alpha }_{metal}-9.45 \\ Wilk{s}^{\prime }\lambda \text{=0}.\text{42},F(\text{2},\text{14})\text{=9}.\text{53},P<\text{0}.\text{002};CanonicalR\text{=0}.\text{76},{\chi }^{\text{2}}\text{=} \\ 12.03,DF=2,p<0.002;ROCED=0.48;ROCFIT=1.21,MC{C}_{Train}= \\ 0.86,MC{C}_{Test}=0.77,AURO{C}_{Train}=0.95,AURC{O}_{Test}=0.90,Cohe{n}^{\prime }s{\kappa }_{\text{Train}}\text{=} \\ 0.85,Cohe{n}^{\prime }s{\kappa }_{\text{Test}}\text{=0}.\text{75},G-mean{s}_{Train}\text{=0}.\text{89},G-mean{s}_{Test}\text{=0}.\text{89} \end{gathered}$$ (1)

The Wilks’ λ value for the LDA equation is 0.42 and the computed values of AUROC for the training and test sets were 0.95 and 0.90, respectively which are much higher than the stipulated threshold value of 0.5. This strongly supports the reliability of the developed discrimination model. The parameter ROCED bearing a value of 0.48 signifies a model closer to the perfect classifier. The LDA model reported highly acceptable values for the MCC, G-_means, and Cohen’s κ for both the training and test sets which strongly account for goodness-of-fit and predictive nature of the model.

The PDD graphs showed that the maximum of the E_H (expectancy to get higher toxic MONPs) and E_L (expectancy to get lower toxic MONPs) values are scattered on different sides of % probability activity (PA) = 50%. As per the analysis of Fig. 3 (Top left), only 10% E_H can be seen in the region of E_L, while no overlapping is noticed in the region of E_H in case of the training set. On the contrary, no overlapping is detected in the region of E_L and while 15% overlapping of E_L is detected in the E_H region (Fig. 3 Top right). Both PDD figures successfully depicted the acceptable discrimination of higher and lower toxic MONPs through the developed model with 90% and 85% accuracy for training and test sets, respectively.

Fig. 3.

(Top) PDD for the training set and the test set; (Middle) Contribution plot for indices to the discriminant functions for higher and lower toxic groups (left) and Euclidean distance-based AD plot (right); and (Bottom) ROC curve of training and test sets where red dots and blue line define fitted ROC curve and grey line signifies random guess. (For interpretation of the references to colour in this figure legend, the reader is referred to the Web version of this article.)

The contribution plot (Fig. 3, Middle left) was developed employing the two modeled descriptors of equation (1) by considering the product of their average values with their corresponding coefficients. Investigating the contribution plot, we can conclude that ε_oxy is evolved as the most significant feature to discriminate higher and lower toxic MONPs with a positive contribution to the equation. The ε_oxy signifies the electronegativity count of oxygen of each MONP. While second feature Σα_metal related to the core environment of metal defined by the ratio of the number of core electrons to the number of valence electrons along with the number of metals present in a specific MONP. Although the discriminating capability between higher and lower toxic MONPs of this feature is quite low but has a positive contribution towards the toxicity of MONPs. Interestingly, with the increased value of both features, the toxicity increases. Thus, to have lower toxicity, MONPs should have a lower value of ε_oxy and Σα_metal. The Euclidean distance-based AD plot confirmed that there is no outlier in the test set compounds and the prediction of all test set MONPs are reliable (Fig. 3, Middle left). The ROC curves for training and test were illustrated in Fig. 3 (Bottom).

3.2.3. Proposed mechanisms of toxic potency in E. coli

Based on the in vitro study outcomes from our earlier study (Puzyn et al., 2011) and the present one, the resultant cytotoxicity data of 25 MONPs showed that only eight MONPs (Er₂O₃, Gd₂O₃, CeO₂, CoO, Co₂O₃, CuO, NiO, and ZnO) were highly toxic to E. coli after 2 h exposure under dark conditions. Previous studies illustrated that both the lanthanide oxide MONPs, Er₂O_3, and Gd₂O₃ showed high toxicity to algal species like Desmodesmus subspicatus and Raphidocelis subcapitata (Devkova et al., 2017). Similarly, Er₂O₃ MONP was highly toxic to zebrafish embryos at a concentration of 50 ppm, showing significant mortality and morphological malformations (Harper et al., 2008). The toxicity of Gd₂O₃ MONP is primarily attributed to the release of Gd³⁺ ion, likewise, it is well established that the toxicity of CuO and ZnO MONPs is mainly due to the metal ion release (Wang et al., 2016). The toxicity of cobalt based MONPs is an intrinsic property and is not related to metal ion release and findings are in line with the previous studies (Wang et al., 2016) The toxicity from the other two lanthanides MONPs, CeO₂ showed toxicity to the higher side for E. coli which is strongly supported by previous studies where researchers found that CeO₂ helps to increase oxidative stress in human bronchial epithelial cells (Eom and Choi, 2009) along with inflammation, and DNA damage in human alveolar epithelial and macrophage cell lines (Lanone et al. (2009); and pulmonary inflammation and alveolar macrophage functional change in rats (Ma et al., 2011). The toxicity of lanthanide MONP is increasing to E. coli along with movement towards the right site of the periodic table and the trend is quite clear [La₂O₃<CeO₂<Gd₂O₃<Er₂O₃]. The low toxicity or no toxicity of WO₃ and Fe₃O₄ was reported in previous studies including (Firouzi et al., 2017; Han et al., 2019). WO₃ is widely used as a visible light photocatalyst (Dong et al., 2017). Due to the low toxicity and high biocompatibility of Fe₃O_4, this MONP has promising biomedical applications which are also confirmed by our in vitro study and previous literature (Kaweeteerawat et al., 2015).

Analyzing the LDA model, one can interpret that electronegativity count of oxygen and the ratio of the number of core electrons to the number of valence electrons along with the number of metal are the significant features to classify and discriminate the toxicity of MONPs to E. coli. A higher value for both features is the resultant of MONPs toxicity to E. coli. The ratio of core electrons to the valence electron is higher for Er₂O₃ and Gd₂O₃ compared to La₂O₃ and other MONPs which is the primary reason for the highest toxicity along with the medium value of oxygen electronegativity which is responsible for detachment of electrons during toxicity. While the core and valence electron ratio is low for CeO₂, but high oxygen electronegativity makes it a toxic MONP. A similar trend is also observed for CoO, ZnO, NiO, and CuO. Again, medium to lower values (lower value of ε_oxy is affecting up to manifold than Σα_metal) for both features are responsible for lower toxic effects of In₂O₃, Mn₂O₃, WO₃, V₂O₃, Bi₂O₃, Sb₂O₃, Al₂O₃, Co₃O₄ and Fe₃O₄ to E. coli. Interestingly, both properties including the oxygen electronegativity count and the ratio of core and valence electron in a MONP strongly support the detachment of metal cations, oxygen anions and electrons for the oxidative stress, damage to DNA and inflammation to E. coli. The identified mechanism for toxicity of 25 MONPs strongly supports the previously suggested toxicity mechanism for 17 MONPs (Gajewicz, 2012; Kar et al., 2014b, 2016; Puzyn et al., 2011) which signifies that the earlier toxicity mechanism prediction was correct as well as with the introduction of eight diverse MONPs in the modeling, the mechanism and interpretation are following the same trends. The toxicity modeling for a large number of MONPs, as in present study helps to build up the confidence of future prediction of new and untested MONPs.

4. Conclusion

One of the largest databases of in vitro cytotoxicity with 25 MONPs was generated by maintaining the same experimental conditions and bioassay protocols using E. coli as a test model. The experimental toxicity data were further utilized to develop seven classification-based models using the periodic table descriptors. The major findings from the in vitro and in silico studies were summarized below:

• Among the 25 tested MONPs, only eight (Er₂O₃, Gd₂O₃, CeO₂, CoO, Co₂O₃, CuO, NiO, and ZnO) were found to be highly toxic to E. coli.

• Three newly introduced lanthanides MONPs [CeO₂ (EC₅₀-64.050 ppm), Gd₂O₃ (EC₅₀-1.112 ppm) and Er₂O₃ (EC₅₀-0.033 ppm)] were identified as the highly toxic ones among all experimentally tested MONPs, while the toxicity of La₂O₃ was found to be on the lower side compared to the other three oxides. After analysis of the toxicity data for lanthanide MONP, one can conclude that the toxicity of these MONPs is increasing towards the right direction of the periodic table, along with the presence of a higher number of electrons. This hypothesis supports the transfer of electrons between the biosystem and the MONPs followed by a free radical generation that contributes to oxidative stress.

• The lanthanides toxicity trend is also observed for MONPs present under Period 4 of the periodic table. Once one moves towards a higher group (4e12), the toxicity of MONPs gradually increases, except iron MONPs. While the trend is not true for Period 5 of the periodic table.

• Out of seven classification-based models, the LDA model evolved as the best predictive model after being tested with stringent qualitative parameters and classification-based metrics.

• The results of modeling based on the LDA model revealed that electronegativity and reductive property of MONPs are responsible for detachment of metal cations, oxygen anions, and electrons leading to an oxidative stress that could cause DNA damage and inflammation in E. coli.

HIGHLIGHTS.

• A cytotoxicity database of 25 metal oxide nanoparticles to E. coli was prepared.

• In vitro analysis identified that lanthanide based MONPs were highly toxic ones.

• Classification based QSAR and machine learning approaches were used to develop in silico models.

• The core electron environment of metal and electronegativity count of oxygen are responsible for cytotoxicity.