Final Report Summary - MOD-ENP-TOX (Modeling Assays Platform "MAP" for hazard ranking of engineered nanoparticles (ENPs))
Executive Summary:
This project aimed to develop a novel and generic modelling assay platform, which can be used as a risk indicator tool to predict the toxicity of metal-based NPs. This platform was based on in-depth analysis of two MeNPs, but which can be further developed to screen the toxicity of a large number ENPs.
Firstly, an extensive database was constructed which includes both well characterized NP and their interaction with biological systems. The database was populated with experimental data from peer-reviewed papers on the in vitro toxicity of spherical amorphous silica NPs, spherical NPs with an amorphous silica shell and zinc oxide NP. The following types of endpoints were included: viability, genotoxicity, inflammation, oxidative stress and apoptosis/necrosis. The database is populated with data of 70 (for silica) and 36 (for ZnO) research papers retrieved for curation. The format of the database is compliant with the standard format ISA-TAB-Nano to make exchange with other researcher possible. Initial analysis of the database revealed the PCCs that are suitable as descriptors for modelling toxicity. With these descriptors a supervised classifier (decision tree) was build.
In a second stage computational mechanistic models were used to unravel the cellular processes underlying NPs toxicity. Cellular process including mitochondrial function, genotoxicity and DNA transcription regulation were analyzed. A model of mitochondrial function was developed comprising the key components of the respiratory chain, ATP production, transporters and free radical production. The model was adapted and calibrated to take into account the effect of some NPs on ROS production. The model on genotoxicity is focussed on the p53/Mdm2 circuit; this circuit plays an important role in cell cycle arrest, DNA repair and apoptosis. Additionally, we developed a kinetic and dynamic model for nanomaterials. This will allow us to understand/predict the faith of NPs in biological systems. A particle size-dependent kinetic model was developed for ZnO NPs in mice. The model will potentially predict dynamic bio-distribution of NPs. The key PCC including partition coefficient and metabolic/elimination rate, were based on published data studying ZnO NPs and Zn+ ions in different tissues of mice. Time-dependent partition coefficients were used and metabolic/elimination rates to determine the adequate kinetic model. The predictability of this model is assessed by calculating the mean absolute percentage error.
The classifier model was validated with different settings and they confirmed the validity of the approach (proof of principle) and the expected impact of the parameters on the classification outcome. The Random Forest proved to be a suitable classification method for NP toxicity assignment. The toxicity grouping created from parameter settings and the viability assays exported from the database strongly depends on the curated assays and measurements added to the database. The experimental setup and the choice of NP concentration series also have a high impact on the toxicity assignment.
The framework presented here operates with a set of pre-defined rules and user-defined thresholds that allow for different levels of stringency, safety margins, and required supporting evidence for categorization. The framework can be used to test the performance of classification under different assumptions and, more importantly, can be easily adapted to a growing and hopefully more and more standardized data basis of NP properties and toxicities.
Project Context and Objectives:
Project context
Our understanding of the biological activity and toxicity of engineered nanoparticles (ENPs) is still incomplete. Many fine particles generally considered “nuisance dusts” are likely to acquire unique surface properties when engineered to nanosize. A growing body of evidence concerns their fate and transport in different environmental matrices. To address the alarming issues about the toxicity of ENPs, there is an urgent need of rational and generic toxicity assessment tools that allow combination of different strategies. In particular, the integration of novel computational models that would reduce reliance on animal testing.
This project aims to develop a novel and generic “Modeling Assay Platform” (MAP) which can be used as a « Risk Indicator » tool to predict the toxicity of metal-based NPs (MeNPs). To demonstrate the feasibility of a MAP prototype, a shortlist of two metal-based ENPs of greatest commercial potential and highest exposure to human and environment has been selected. The selected MeNPs are depicted from the list of nanomaterials in phase one of the OECD testing program and are silica and ZnO NP.
Project objectives
The project is a multidisciplinary project, aiming to accomplish a development a novel and generic “Modelling Assay Platform (MAP)” which can be used as a « Risk Indicator » tool to predict the toxicity of MeNPs; The projects also aims to demonstrate the feasibility of a MAP prototype, based on in-depth analysis of two MeNPs, which can be further developed to screen the toxicity of a large number ENPs. Based on the concept of integrating different testing strategies, the proposed generic MAP combines two main and complimentary paradigms: (1) a novel Computational Modeling Package (CMP) based on structural, mechanistic, as well as kinetic modeling tools and (2) an innovative screening (IS) strategy that allows performing multiplexed streamlined and predictive assays for calibration, refinement and validation of the computed models.
More specifically, MOD-ENP-TOX activities aim to fulfill six scientific and technology objectives:
Objective 1: Identify structural and physico-chemical characteristics that drive the toxicity of MeNPs.
The main objective is to perform a comprehensive survey of existing toxicity studies on two
MeNPs and construct a sound joined database (dataset) based on well-selected sources. Important consortia and toxicology labs actively involved in the field will be contacted to acquire access to their database(s). We will also make use of public sources, and available database from Health and Environmental Services. In these databases, relevant physico-chemical characteristics (PCCs) describing the material used and the measurable biological endpoints assessed will be combined. Based on the compiled “joined database”, classification models (Classifiers) will be generated based on key descriptors in order to classify the selected MeNPs into groups of similar toxicity patterns.
Objective 2: Identify the relation of MeNPs descriptors to toxicological endpoints and classify a training set of MeNPs into groups of regular toxicity patterns.
Based on this database, machine learning algorithms and multivariate analysis will be used to develop first dendrograms of hierarchical clustering of MeNPs and then a classifier model to define groups of ENPs having regular toxicological profiles. A successful classification will be achieved through identification of the different key descriptors (structural and PCCs) of MeNPs known to induce cellular toxicity. These descriptors will be suitably correlated to identify toxicological endpoints and used as readouts for building the computational models.
Objective 3: Development of a Computational Mechanistic Package (CMP) to predict MeNP toxicity.
A computational modeling package will be constructed to predict MeNPs toxicity. This consists in the implementation of an in silico model that integrates key cellular processes
underlying MeNP toxicity, including oxidative stress (ROS formation), mitochondria metabolism and bioenergetics, as well as various cellular cascades involved in cell permeability, DNA damage, and inflammation.
Objective 4: Develop a kinetic and dynamic model based on PBPK to predict MeNPs toxicity in human health.
In-vivo toxicity data from the database will be used to construct a kinetic PBPK model to
simulate the dynamics of ADME from different routes of MeNPs exposures. Such a PBPK model is significantly useful to evaluate kinetic of NPs mobility, disposition and associations between exposures and biomarker measurements in different physiologically relevant compartments (i.e. blood, urine, kidneys, liver, brain etc..).
Objective 5: Establish and utilize in-vitro & in-vivo innovative screening toxicity assays to calibrate and validate the CMP and the PBPK model.
We will design appropriate in vitro/in vivo assay models to generate experimental data needed to validate the computational models and perform a thorough characterization of the MeNPs used in the in vitro/in vivo assays. These data will be used to validate the computational models which will be developed.
Objective 6: Integrate all the testing strategies into a single modeling platform and demonstrate the feasibility of a MAP prototype.
The final step will be to integrate the constructed models into a platform, and to define an “overall estimate for toxicity / hazard potential”.
Project Results:
The MOD-ENP-TOX project developed a novel and generic modelling assay platform, which can be used as a risk indicator tool to predict the toxicity of nanoparticles (NP). This platform was based on in-depth analysis of two metal-based NPs, but which can be further developed to screen the toxicity of a large number NPs.
Below we will give a concise overview of the work done throughout the whole project. Firstly, an extensive database was constructed which includes both well characterized NP and their interaction with biological systems. The database was fed with data from literature, complemented with own experimental data. In parallel, different computational models were designed to unravel the cellular processes underlying NPs toxicity. These models included mechanistic, as well as kinetic and dynamic computational models. A new modelling platform was designed. Finally, the platform was tested and validated with the data extracted from the database.
1. Construction of a database and identification of key descriptors of the training set of metal-NPs and their classification into groups
Initially KU Leuven chose to populate the database with existing in vitro toxicological data
from the following sources: internal data of the S2Nano-project (the Belgian program “Science for sustainable development”); the ENPRA-project (a European project to develop and implement a novel integrated approach for engineered NP risk assessment); and literature data of peer reviewed papers. After investigation of the different data sources, it was decided to use data from published peer reviewed papers on the in vitro toxicity of amorphous silica nanoparticles and supplement these data with the internal toxicity data from silica nanoparticles of the he S2Nano-project. Until date, data of in vitro experiments on the most common investigated toxicological endpoints for NPs were collected; cell viability, apoptosis/necrosis, genotoxicity, oxidative stress and inflammation.
Different possibilities of how to store the curated data sets were reviewed. A human readable format was favoured to store and exchange data or datasets. Genedata introduced two human-readable database formats, the first GeneData Arrays (GDA) in order to store assay endpoints data and the second one GeneData Columns (GDC) to store the PCCs. These formats allow to be read by other software due to its logical structure. The GDA/GDC format can be used to import data into Genedata Analyst, software for data integration and analysis. Genedata Analyst will be used to perform tasks such as classification but also integration and grouping of data sets from different labs and different experimental setups. The consortium participated in the US Nano Working Group to discuss the new ISA-TAB-Nano format, which was published during the first year of the project. It was decided that the usage of GDA/GDC format supporting ISA-TAB-Nano nomenclature was the most efficient format to perform the upcoming data analysis steps and still being able to exchange data in ISA-TAB-Nano once a parser is written to convert the different file formats. The GDA/GDC file format was used to curate the data of publications which have been selected as the first approach to perform a toxicity classification of NP. After entering the data, the GDC/GDA files were reviewed and nomenclature and errors in these data could be easily identified using Genedata Analyst by checking for inconsistencies (e.g. spelling errors and duplicate data entries).
A detailed quality evaluation is needed both for the measured physico-chemical characteristics (PCCs) of the NP and their interaction with biological systems. As agreed with other consortia, ten PCCs that might determine the toxicity of NPs were included in the database: size distribution, agglomeration state, shape, crystal structure, chemical composition, surface area, surface chemistry, surface charge, solubility and porosity. Within the MOD-ENP-TOX project, these ten PCCs were divided in three groups of priority: obligatory (shape, size, composition and crystallinity), optional but very important (surface area, surface chemistry and surface potential) and optional (solubility, porosity and agglomeration/aggregation). In addition, also the reliability of the characterization method used was introduced into the GDC. A supplementary purity score and also an overall score for the amount of characteristics
specified or measured in the article were added. As there are no international standards for NP testing it is difficult to score the reliability of the assays used. Data evaluation is based on: the amount of information that is provided about the assay, the use of positive and negative controls and whether are not possible interaction of the assays with the NP were considered.
The development of a tool to classify the toxicity of NP requires a sufficient number of data points in order to define the toxicity of a NP based on the respective assay readout. A toxicity classification based on the PCCs needs a reasonable amount of NP which populate groups, toxic and a non-toxic. The cytotoxicity assays were chosen as a first group of assay readouts which were integrated as they contain different NP concentration levels which can be used as an indicator of level of toxicity. A viability drop of more than 25% was defined if a NP is regarded as toxic. The concentration level of the respective NP defines the level of toxicity. Therefore, four different toxicity levels were defined:
- Toxicity Level 0 (Not proven toxic (NPT)): a viability assay readout based on NP concentration >250 μg /ml annotates the NP as not proven toxic
- Toxicity Level 1 (Low Toxicity): a viability assay readout based on NP concentration >150-250 μg /ml annotates the NP as low toxic
- Toxicity Level 2 (Medium Toxicity): a viability assay readout based on NP concentration >75-150 μg /ml annotates the NP as medium toxic
- Toxicity Level 3 (High Toxicity): a viability assay readout based on NP concentration ≤75 μg /ml annotates the NP as highly toxic
The data subtracted from peer review papers as described above were generated in different labs without an underlying standard operating procedure. The heterogeneity of the data from the same viability assay readout is caused by several factors such as NP concentration units (e.g. mass per volume, mass per surface area, particle surface area per volume ...) cell type or contact time. Within the MOD-ENP-TOX consortium, a data integration strategy was setup on how to make these data sets comparable (if conversion factors are reported in the papers) and how to integrate the results. Although the importance of these features is not questioned, setting the focus on the main factors such as NP concentration was essential in order to perform a first toxicity grouping approach. After the available NPs have been assigned, a toxicity level based on their assay readouts, the impact of the PCCs of the NPs was analyzed. A Principal Components Analysis revealed that the PCC size is probably the best characteristic to differentiate between toxic and non-toxic NP. The individual PCCs were analyzed, but revealed no significant effect, except for the size. The effect of regrouping of the four toxicity levels (assigning a low toxic NP to the non-toxic group and thus reducing the number of different groups) was analyzed. Several classification terms were applied, the Decision Tree proved to be best suited for classification, which was verified by Genedata using a Leave-One-Out cross validation.
Data on the in vitro toxicity of amorphous silica NPs, particles with an amorphous silica shell and zinc oxide NP were collected from peer reviewed papers. Papers investigating particles with a defined shape (aspect ratio (AR) <3), composition (silica, silica shell or zinc), crystallinity (amorphous) and primary size were included. Apart from the obligatory characteristics (primary size, shape, composition and crystallinity), other physicochemical determinants of toxicity were also recorded: the surface charge, surface area, agglomeration/aggregation status, solubility and porosity of the particles. Five types of toxicological endpoints were included in the database: cell viability, apoptosis/necrosis, genotoxicity, oxidative stress and pro-inflammation. For data regarding in vivo toxicity of silica NP additionally included studies on the effect of the particles on mice or rat (experimental in vivo studies), and information on bio distribution, growth and consumption, blood cell count, energy metabolism, genotoxicity, pro-inflammation, oxidative stress or apoptosis/necrosis (see Figure 1) After extensive filtering based on the completeness of the data, only 5 papers were retrieved to be included in the database.
From the initial Medline search, which revealed 450 relevant papers, 70 papers were retrieved to construct the database, describing 147 different silica NP (or NP with a silica shell). For ZnO NP, a list of 36 papers was retrieved to construct the database, describing 70 different ZnO NPs (or NPs with a zinc oxide shell) and their characteristics in 229 different assays. Data were manually curated from text, tables and figures. To make the database exchangeable, data are stored in ISA-TAB-Nano, a standard format for sharing nanomaterial research data.
The database was further enriched by generating data from own experiments and feeding them into the database, in order to allow for validation of the computational models. Since the main toxicity pathway investigated in the database and the computational models to predict the toxicity of amorphous SiO2 and ZnO NPs is oxidative stress, both in vitro and in vivo experiments focused on the general toxicity and oxidative stress as a pathway underlying this general toxicity.
In vitro testing was done using the cell types which best illustrate the major ways NPs may enter and distribute in the human body such as epithelial cells and macrophages (inhalation), endothelial cells (blood circulation). First, the effects of different concentrations of silica NP were measured in viability assays (in human bronchial epithelial cells and human blood monocytes) via WST-1 reduction assays and the LDH release assays. The smaller the amorphous silica NPs the more cytotoxic they were for the epithelial cells and the monocytes. The cytotoxicity of the purpose made SiO2-10 and commercial available ZnO NP was tested on mouse lung epithelium cells. The MTT reduction assay was used to measure viability of the cells after 48 hours of exposure to different concentrations of the NPs. Again a dose-dependent response was observed after SiO2-10 and 10 nm ZnO NPs exposure.
To further investigate the mechanism responsible for the loss in viability of the cells, oxidative stress caused by the NPs was measured. Mouse lung epithelium cells were exposed to a sub-toxic concentration of ZnO NP for different incubation periods. After exposure reactive oxygen species (ROS) production was measured with the DCFH-DA assay. The intracellular ROS-production of the epithelium cells exposed to ZnO NPs for 6 and 16 hours was significantly higher than in cells exposed to culture medium. Since oxidative stress is closely related to inflammatory response by activating e.g. the nuclear factor-kappa B (Nf-κB) and p53 signalling pathway, the time- and dose-dependent effect of ZnO and SiO2-10 NPs on IKB-α and p53 proteins in these cells was investigated with western blot.
All these experiments allow quantifying the cellular response at different time-points. However, cell culture assays cannot reliable model the higher level interactions which are presented in living organism. Therefore, in vivo studies were needed.
Firstly, zebrafish embryo/larvae were exposed to SiO2-10 NPs, ZnO NPs, Zn(NO3)2 or vehicle in their medium. At different time point, they were washed and the amount of Si or Zn that they incorporated by the zebrafish was determined. No significant differences were observed for the different treatment groups. However, the results suggested an active uptake of the ZnO NPs and this in contrast with the Zn ions which are not taken up by the larvae. Also silica seems to be taken up by both the embryos and larvae.
Glutathione (GSH) is an important anti-oxidant of the cell, and is oxidized to GSSG when there is oxidative stress caused by reactive oxygen species. The ratio between GSSG and GSH is used as a measure for cellular toxicity. For the four different treatment groups the amount of GSSG and GSH per mg proteins was measured in the zebrafish 6 to 120 hours post fertilization. No significant differences were observed between the treatment groups.
Secondly, the bio-distribution of the NP in mice was investigated, as well as the effect on oxidative stress and inflammation pathways. Concentration of Zn and Si were measured at different time-points after the intravenous dosage of vehicle, 10 nm ZnO NPs, SiO2-10 NPs or Zn(NO3)2. The blood, brain, lung liver, kidney and spleen of the mice were analyzed with ICP-MS. No statistical differences were observed between different time-points and treatment groups after non-parametric tests. Reduced and oxidized levels of glutathione were also measured in these organs and at the same time-points. No statistically significant differences can be seen between different treatment groups due to the high variance within these groups. IKB-α and p53 concentrations were also measured in the liver, spleen, brain and the lung of the mice at the different time-points.
In a final stage, a full set of characteristics was reported for each NP used. The characterization is done in house to ensure that the same protocols were used for all NPs, and that the NPs have the desired PCCs analyzed. Since most techniques have some drawbacks, an array of techniques is needed to fully characterize the NP. The NPs were characterized with SAXS, ZP, DLS, TEM for size analysis. The PCCs determining the toxicity of NPs are: chemical composition and purity of the NPs measured by inductively coupled plasma mass spectrometry (ICP-MS), solubility measured by dialysis experiments, crystal structure measured by transmission electron microscopy (TEM), particle size distribution measured by TEM, small angle x-ray scattering (SAXS), dynamic light scattering (DLS), shape measured by TEM and SAXS and agglomeration and/or aggregation measured by TEM, SAXS and DLS. Surface charge was measured by zeta-potential measurement, surface chemistry, measured by Fourier transform infrared spectroscopy (FTIR) and surface loading, measured by thermogravimetric analysis (TGA).
Besides the measurements of the PCCs, their stability was tested in Danieau’s medium, since this medium was used to maintain zebrafish larvae which were used in the in vitro experiments as described above. This experiment ensured stable concentrations of the NPs throughout the duration of the in vitro experiments. Therefore, we monitored the NPs before and after they were exposed to the larvae and Danieau’s medium with both SAXS and DLS. The size monitored with DLS showed no difference in size before and after the exposure. Furthermore, the concentration was monitored with SAXS. These results showed that the concentration did not deviate after exposure.
Taken together, all these experimental data were added to the database and used to validate the computational models, as described below.
2. Development of a Computational Mechanistic Package (CMP) and a physiologically-based pharmacokinetic (PBPK) model to predict the toxicity of MeNPs in a quantitative way.
In parallel, three different computational models were designed to predict the toxicity of NP, which were fed with the data from the database described above.
Development of a computational model of mitochondria
A computational model of mitochondria was developed, which took into account appropriately balanced mass, charge, and free energy of various ions and electrons as well as the transduction systems. The model was based on a set of previously published data. Both the experimental data and computational models were previously developed for cardiac mitochondria.
The basic components we included in the new mitochondrial model are:
- The reactions of the mitochondrial respiratory chain for complexes I, III, and IV of the electron transport system and ATP synthesis for the F1F0 ATPase
- The substrate transporters, including adenine nucleotide translocase and the phosphate–hydrogen co-transporter
- Cation fluxes across the inner membrane, including fluxes through the K+/H+ antiporter and passive H+ and K+ permeation
- In addition, we integrated in our new mitochondrial model the production of the free radical superoxide (O2-), which is dependent on the NADH concentration and the mitochondria membrane potential (ΔΨm). Superoxide is transported from the matrix to the cytosol via a transporter called IMAC. The enzyme superoxide dismutase (SOD) catalyses the dismutation of superoxide (O2-) into hydrogen peroxide H2O2. H2O2 is highly toxic, decomposes with formation of free radicals, but it is rapidly neutralized and transformed into water by endogenous efficient defence molecules, e.g. the antioxidant glutathione (GSH) and the enzyme catalase (CAT).
It is important to notice that all the models published in the literature only considered one
component or a very small set of components of the mitochondria, and none previously integrated the different mitochondrial components together. By contrast, in our mitochondrial model, we integrated all the components needed for the respiratory chain and ATP production, the transporters, the mitochondrial permeability potential as well free radical production. In addition, our model is the only model, which integrates the effect of metal NP on the mitochondrial permeability potential (↓ΔΨm) and the oxidative stress chain (↓catalase).
Development of a computational model of p53 activation dynamics
Two deterministic models of the p53/Mdm2 circuitry were developed, based on previously published models reflecting biological reality with a minimum level of detail. We considered two different mechanisms for the DNA damage response:
- ARF activation, followed by sequestering of Mdm2
- ATM activation followed by phosphorylation of Mdm2 and p53
To simulate the effect of a gradual increase in DNA damage induced by MeNPs, we introduced a species called ROS (reactive oxygen species) into the model. The model predictions regarding the oscillation period are in agreement with experimental data. Under conditions of irradiation, this model generally produced one large initial peak of p53 after DNA damage, followed by irregular oscillations in which p53 levels remained relatively high.
Our model integrating the above biomarkers and their connections is unique. Indeed, ROS
input to the ARF and ATM was not published, and the in silico connection between mitochondria ROS production and p53 activation was not yet published. The dynamic of p53 oscillations remains unclear in experimental studies and oscillation is very difficult to measure experimentally. The in silico p53 model proposes a rational explanation for p53 oscillation dynamics: it depends on two stimulation modes dependent either on ATM or ARF. Such an in silico model may be useful for discovery of strategies (target and compound) aiming at modulating such oscillations or for supportive their use as a biomarker: p53 could become a new readout (biomarker) for MNP toxicity testing. Likewise, ARF and ATM models are generic models, which could be used for oncology research and for safety screening (effects of MNPs alone or in association with other molecules).
Development of a computational model of NFκB
Based on simplifying reductions of the IκB–NFκB signalling, a computational model was developed, describing the temporal control of NFκB activation by the coordinated degradation and synthesis of three IκB proteins isoforms (IκBα, IκBβ, and IκBε). The model demonstrated that IκBα is responsible for the strong negative feedback that allows for a fast turn-off of the NFκB response, whereas IκBβ and IκBε function to reduce the system's oscillatory potential and stabilize NFκB responses during longer stimulations.
NFκB is a protein complex that controls transcription of DNA. Based on simplifying reductions of the IκB–NFκB signalling, we developed a computational model that describes the temporal control of NFκB activation by the coordinated degradation and synthesis of IκB proteins isoforms (IκBα, IκBβ, and IκBε). The model demonstrates that IκBα is responsible for strong negative feedback that allows for a fast turn-off of the NFκB response, whereas IκBβ and IκBε function to reduce the system's oscillatory potential and stabilize NFκB responses during longer stimulations. Varied by two orders of magnitude on either side of the standard rate value of the three rates constant, IκBα transcription rate constant, IκBα nuclear import and rate constant for IKK degradation is a concrete example of how to use modelling to understand oscillation and the underlying mechanisms that delay them and could deregulate the pathway and induce toxicity and cell death.
Of importance, the results generated in silico, using the mitochondrial model, were validated against data of published experiments. To validate the model into a broader context and to become reliable to predict effect over time using different sizes of particles, the computational model needed additional test results of in vitro experiments. The limited (but crucial) set of tests were setup, including effects of MeNPs on diverse molecular targets implicated in cytotoxicity (or cyto-survival), such as inhibition of glutathione or blockade of complex I, in addition to downstream molecular targets, p53 and NFκB.
Development of a physiologically based pharmacokinetic model to predict MeNPs effects on human and demonstration of the feasibility of the model using subsets of MeNPs
A particle size-dependent PBPK model for ZnO NPs in vivo was developed. The model predicted dynamic bio-distribution of MeNPs. We estimated key physicochemical parameters of partition coefficient and metabolic/elimination rate based on our previously published data quantifying the biodistributions of 10 and 71 nm ZnO NPs and Zn(NO3)2 in different tissues of mice. Time-dependent partition coefficients were used and metabolic/elimination rates to determine the adequate PBPK model. The predictability of this model was assessed by calculating the mean absolute percentage error.
The PBPK models provides a potential effective tool to estimate the time course of chemical accumulation in target tissues of organism and can be incorporated into a quantitative risk assessment framework. Therefore, the development of a MeNPs PBPK model with the highly adequate predictive power is urgently needed to gain insights into underlying mechanical processes, hypothesis testing, and to guide experimental design. The advantages of a nanometal PBPK model not only can reduce animal testing and cost, but also can simulate and predict bio-distribution in human and human response to NP.
3. Validation and calibration of the computed models by quantitative measurements of toxicity in in vitro and in vivo assays and integration of all testing strategies into a modelling platform
Due to an unforeseen termination of participation of one of the consortium members, the original planning to validate and interconnect the three newly developed computational models needed to be revised. As a consequence, the core of the project was focused on one classifier model. The classifier model was validated on using experimental data in the database from 4 amorphous silica NP and its viability assays. As the advanced classification tool allowed the user to influence the toxicity group assignment of the NP used for the training of the classifiers by specific input parameters, the effect of the chosen parameters on the classification outcome was investigated. For the final validation of the classifier using experimental in vitro data from viability LDH release assay and WST-1 reduction assay the different input parameters were chosen to compare the results.
It was concluded that the new input parameters proved to have a high impact on the assignment of NP to the toxicity groups toxic and non-toxic. These input parameters were Viability Threshold, NP Toxicity Concentration Threshold, NP Non-Toxicity Concentration Threshold, Minimum Number of Confirming Measurements. The evaluation setup of comparing 26 runs with different settings confirmed the validity of the approach (proof of principle) and the expected impact of the parameters on the classification outcome. The Random Forest proved to be a suitable classification method for NP toxicity assignment. The classifier performance was shown to highly depend on the settings which determine the initial toxicity group assignment of the NPs and has been confirmed in most runs to deliver AUCs of ROCs of about 0.75. With optimal settings an area under the curve (AUC) of 0.87 was achieved, which indicates a fairly good prediction accuracy. A schematic overview is given in Figure 2.
Therefore, a number of improvements have been included in the new version of the model, since a small change in the viability threshold could have a strong effect on the population of toxic and non-toxic groups. The same applies to the NP concentration threshold at which a positive (toxic) assay output refers to a toxic annotation of the NP. To our best knowledge, there are no standard values when a viability assay is regarded to have a positive outcome. For that reason, the classification model has been extended to allow the user to create individual toxicity/non-toxicity NP groups depending on customized thresholds for both the viability and the NP concentration. Another feature has been added to the classification model to define the number of minimum data points necessary to label a NP as toxic or non-toxic. Although the data points from a variety of viability assay data from a multitude of publications are computed based on replicates and the potential of outliers should be excluded, it may still be possible that an error persists in the database. Most viability experiments are thoroughly set up with several NP concentrations in order to get more than just one measurement. However, increasing the minimum number of data points can significantly reduce the population of the toxic and non-toxic NP groups especially for values > 2.
The differentiation of four toxicity groups (described above, namely not proven toxic, low, medium and high toxicity) was based on the assumption that enough NP data is available to populate the groups for subsequent classification. As this assumption did not hold, we decided to differentiate only between two groups to provide more robust classifications using one concentration threshold 75 µg/ml.
The drawback of this setup is that such a threshold is crucial and very similar concentration levels can still result in two opposite assignments (e.g. 75 µg/ml and 76 µg/ml). Classification and cross-validation tests with regrouping of toxicity levels and excluding NPs assigned to a group of “insecurity”, however, seemed to perform better (data not shown). Toxicity groups NPT and Low Toxicity have been merged to non-toxic, High toxicity has been assigned as toxic. The medium toxicity group has been omitted as the group of “insecurity”.
Based on these results, we decided to develop thresholds for the two groups toxic and non-toxic and added the following assumptions:
Toxic NP Group: Viability < Viability Threshold (default: 75%) and NP Concentration < ConcTox Threshold (default: 75 µg/ml)
Non-Toxicity NP Group: Viability ≥ Viability Threshold (default: 75%) and NP Concentration > ConcNonTox Threshold (default: 75 µg/ml)
These two threshold parameters ConcTox and ConcNonTox are exposed to the user and will have a high impact on the toxicity group assignment. Due to the added functionality which allows for dynamically changing the settings described above, the toxicity assignment has to be performed for every single classification run. Therefore, the extracted data set without prior NP toxicity assignments needs to be available.
Subsequently, the new functionality was evaluated by running two classification evaluations with the extracted viability data set of the database for silica NP with the varying preset parameters of viability threshold, NP toxicity concentration threshold, NP non-toxicity concentration threshold and the minimum number of data points. Both evaluation models confirm the good performance of the Random Forest classification method. While the settings used in Evaluation A (viability threshold of 50% and a NP toxicity concentration threshold of 75 µg/ml) provides a better estimate of toxicity for the underlying data set, Evaluation B (viability threshold of 75% and a NP toxicity concentration threshold of 150 µg/ml) proves that a different choice of parameter settings does not have the same impact on the classification models.
As the Random Forest performed best on the underlying data set, this classifier was used in order to predict toxicity for all NPs to be analyzed. The result of the cross-validation has been added to the toxicity prediction. As the user did not have influence on the underlying data set for the validation, the result image was static and did not change.
Due to our changes and added functionality explained above, the underlying data set is now dynamically created depending on the settings of the user. Also the next step, i.e. the evaluation of the classification performance is now performed dynamically based on the user-defined validation dataset. In each run of the tool the available classification methods are compared on the validation dataset by means of nested cross-validation. The best performing prediction model is then used for classifying the NP under investigation based on the respective PCC measurements. The outcome of the comparative performance evaluation and the classification of the input NP are presented in an HTML report. This report includes a dynamically created image of the averaged ROC curves, which were used for assessing the prediction accuracy of the constructed models, as well as a table specifying the estimated probability that the input NP is toxic and non-toxic, respectively.
Taken together, the advanced classifier proved to cope with different experimental setups and allows to assign toxicity for different thresholds on viability as well as NP concentration thresholds. Although the validation of the classifier was limited to four NP, 2 different assay readouts and 6 different NP concentration steps, the classifier proves to provide comparable toxicity estimates. Both the classifier estimate output and the experimental data give proof that the change of thresholds used to define NP toxicity has a huge impact on the outcome. The classifier model returned an estimate of the toxicity of a NP under investigation based on its PCC’s. The model was trained on data from 116 NP measured in different viability assays.
4. Modeling platform infrastructure
The classifier model is accessible via a web user interface. On the main page a screen is offered to the user at which five parameters (five common particle characteristics) need to be filled out: Diameter – or size (in nm), Specific Surface Area: (m²/g), External Surface Area: (m²/g), Zeta potential in water (mV) and Hydrodynamic diameter in water (nm).
After clicking the button ‘Run script’ an additional tab is opened showing the outcome of this specific simulation. Two different types of outcome data are provided: The statistical probability calculation whether the material (of the simulation) will be toxic (meaning inducing cytotoxicity (reduction of more than 25 %) in vitro after 24 h at a concentration of 75 µg/ml or less). This probability is given in percentage. The performance of the evaluated classifier algorithms is compared using a ROC (receiver operating characteristic) curve. Also the algorithm used to generate the given probability is specified.
Based on the validation data set of toxic and non-toxic NPs, which has been extracted from the database according to user-defined toxicity parameters, the classification performance of five state-of-the-art supervised classification algorithms (Random Forest, K-Nearest Neighbors, Nearest Shrunken Centroids, Weighted Voting and Adaptive Boosting) is determined in a nested cross-validation. The best classifier is selected based in the highest average area under the ROC curve. This classifier is then applied to predict the toxicity of the NP under investigation. The returned prediction score is transformed into a value between 0 and 1 which corresponds to the probability that the input NP is toxic. The outcome of the method comparison and the classification of the input NP are presented as shown in Figure 3.
The modelling platform setup allows access management using a name and password. Results from mechanistic modelling runs were stored and could be reopened using results management infrastructure. Users can register or log in if they already have credentials to run a classification model. In addition, the new platform also allows to store results from the classification model as well as the input parameters. The generated results can easily be downloaded or shared with other users. The info button gives also access to valuable information including when a modelling job has been started and when it has been finished.
Potential Impact:
The MOD-ENP-TOX project aimed to develop a novel and generic modeling assay platform, which can be used as a risk indicator tool to predict the toxicity of metal-based NPs. This platform was based on in-depth analysis of two MeNPs, but which can be further developed to screen the toxicity of a large number ENPs. To illustrate the impact of the MOD-ENP-TOX project on European research and industry, we will describe how the project contributed to the enhanced the current scientific knowledge, but also how the project led to new collaborations and future research projects.
Firstly, an extensive database was constructed which includes both well characterized NP and their interaction with biological systems. The database was populated with experimental data from peer-reviewed papers on the in vitro toxicity of spherical amorphous silica NPs, spherical NPs with an amorphous silica shell and zinc oxide NP. The database is set up following the internationally agreed format ISA-TAB-Nano. This will allow external groups to extract and use data from the database format. Due to this standardized format and high quality (including criteria were used to guarantee the quality of the data in the database), the database is not only useful within the MOD-ENP-TOX project but can also be used by other stakeholders who wish to determine quantitative structure/activity relations for nanomaterials. Therefore, the full database will become public on our server after 1-2 years after our internal exploitation.
In a second stage computational mechanistic models were used to unravel the cellular processes underlying NPs toxicity. The main challenge in predicting toxicity using classifiers based on literature data comes from the fact that assays and readouts used for toxicity assessment as well as determination of PCCs are diverse, and comparable only to a very limited extend. A modelling approach, as presented within MOD-ENP-TOX, has to cope with a sparse data matrix (i.e. the exact same readout is available only for very few nanoparticles) which makes assumptions necessary under which circumstances a result shall be regarded “positive” or “negative”, respectively. The framework presented here operates with a set of pre-defined rules (e.g. a nanoparticle classified as “toxic” in one assay will always be regarded as toxic, even if another assay classified it as “non-toxic”) and user-defined thresholds, that allow for different levels of stringency, safety margins, and required supporting evidence for categorization. The framework can be used to test the performance of classification under different assumptions and, more importantly, can be easily adapted to a growing and hopefully more and more standardized data basis of nanoparticle properties and toxicities.
The consortium members actively participated different European initiatives throughout the project, contributing to the enhancement of the knowledge about risk assessment for the use of nanomaterials. Nanotechnology produces engineered nanomaterials having new or enhanced physico-chemical properties in comparison to their micron-sized counterparts. Some of these properties, like the high surface area to volume ratio, make them potentially dangerous to humans. To promote the development of a new generation of nanomaterials that are safe-by-design, an understanding of the relationship between the structure of the nanomaterial and biological activity is needed. In this context the Quantitative Nanostructure-Toxicity Relationships computational modelling technique is an effective alternative to experimental testing since it enables the prediction of (eco)-toxicological effects based on nanomaterial structure only.
The construction of these models requires the integration of expertise of nanomaterial scientists, (eco)-toxicologists, and modellers from academia, regulatory agencies and industry. Therefore, a network for trans-disciplinary cooperation was needed, which resulted in the COST action entitled ‘Modelling Nanomaterial Toxicity (MODENA).
This COST action was established to promote and to co-ordinate these inter-disciplinary collaborations, with the ultimate aim of producing computational models for engineered nanomaterials. Peter Hoet of the KU Leuven is one of the responsible persons representing Belgium in these activities.
As a scientific outcome of these actions, a new collaboration has been setup between Sabine van Miert (Thomas Moore, Geel Belgium), Peter Hoet and Jean-Pierre Locquet (KU Leuven, Belgium). The database developed by the MOD-ENP-TOX project will be used by the modeling groups of Sabine Van Miert to exploit in more detail PBPK models and classifier models. These efforts will be done outside the contact of the project but will give the data collected in the project an ‘new’ live since the potential of the whole database was not yet fully exploited. Secondly, Sabine van Miert (Thomas Moore, Geel Belgium), Peter Hoet (KU Leuven, Belgium) and Helena Oliveira (University Aveira, Portugal) are building a PBPK model for AgNPs (no MOD-ENP-TOX data).
List of Websites:
http://fys.kuleuven.be/apps/modenptox/
This project aimed to develop a novel and generic modelling assay platform, which can be used as a risk indicator tool to predict the toxicity of metal-based NPs. This platform was based on in-depth analysis of two MeNPs, but which can be further developed to screen the toxicity of a large number ENPs.
Firstly, an extensive database was constructed which includes both well characterized NP and their interaction with biological systems. The database was populated with experimental data from peer-reviewed papers on the in vitro toxicity of spherical amorphous silica NPs, spherical NPs with an amorphous silica shell and zinc oxide NP. The following types of endpoints were included: viability, genotoxicity, inflammation, oxidative stress and apoptosis/necrosis. The database is populated with data of 70 (for silica) and 36 (for ZnO) research papers retrieved for curation. The format of the database is compliant with the standard format ISA-TAB-Nano to make exchange with other researcher possible. Initial analysis of the database revealed the PCCs that are suitable as descriptors for modelling toxicity. With these descriptors a supervised classifier (decision tree) was build.
In a second stage computational mechanistic models were used to unravel the cellular processes underlying NPs toxicity. Cellular process including mitochondrial function, genotoxicity and DNA transcription regulation were analyzed. A model of mitochondrial function was developed comprising the key components of the respiratory chain, ATP production, transporters and free radical production. The model was adapted and calibrated to take into account the effect of some NPs on ROS production. The model on genotoxicity is focussed on the p53/Mdm2 circuit; this circuit plays an important role in cell cycle arrest, DNA repair and apoptosis. Additionally, we developed a kinetic and dynamic model for nanomaterials. This will allow us to understand/predict the faith of NPs in biological systems. A particle size-dependent kinetic model was developed for ZnO NPs in mice. The model will potentially predict dynamic bio-distribution of NPs. The key PCC including partition coefficient and metabolic/elimination rate, were based on published data studying ZnO NPs and Zn+ ions in different tissues of mice. Time-dependent partition coefficients were used and metabolic/elimination rates to determine the adequate kinetic model. The predictability of this model is assessed by calculating the mean absolute percentage error.
The classifier model was validated with different settings and they confirmed the validity of the approach (proof of principle) and the expected impact of the parameters on the classification outcome. The Random Forest proved to be a suitable classification method for NP toxicity assignment. The toxicity grouping created from parameter settings and the viability assays exported from the database strongly depends on the curated assays and measurements added to the database. The experimental setup and the choice of NP concentration series also have a high impact on the toxicity assignment.
The framework presented here operates with a set of pre-defined rules and user-defined thresholds that allow for different levels of stringency, safety margins, and required supporting evidence for categorization. The framework can be used to test the performance of classification under different assumptions and, more importantly, can be easily adapted to a growing and hopefully more and more standardized data basis of NP properties and toxicities.
Project Context and Objectives:
Project context
Our understanding of the biological activity and toxicity of engineered nanoparticles (ENPs) is still incomplete. Many fine particles generally considered “nuisance dusts” are likely to acquire unique surface properties when engineered to nanosize. A growing body of evidence concerns their fate and transport in different environmental matrices. To address the alarming issues about the toxicity of ENPs, there is an urgent need of rational and generic toxicity assessment tools that allow combination of different strategies. In particular, the integration of novel computational models that would reduce reliance on animal testing.
This project aims to develop a novel and generic “Modeling Assay Platform” (MAP) which can be used as a « Risk Indicator » tool to predict the toxicity of metal-based NPs (MeNPs). To demonstrate the feasibility of a MAP prototype, a shortlist of two metal-based ENPs of greatest commercial potential and highest exposure to human and environment has been selected. The selected MeNPs are depicted from the list of nanomaterials in phase one of the OECD testing program and are silica and ZnO NP.
Project objectives
The project is a multidisciplinary project, aiming to accomplish a development a novel and generic “Modelling Assay Platform (MAP)” which can be used as a « Risk Indicator » tool to predict the toxicity of MeNPs; The projects also aims to demonstrate the feasibility of a MAP prototype, based on in-depth analysis of two MeNPs, which can be further developed to screen the toxicity of a large number ENPs. Based on the concept of integrating different testing strategies, the proposed generic MAP combines two main and complimentary paradigms: (1) a novel Computational Modeling Package (CMP) based on structural, mechanistic, as well as kinetic modeling tools and (2) an innovative screening (IS) strategy that allows performing multiplexed streamlined and predictive assays for calibration, refinement and validation of the computed models.
More specifically, MOD-ENP-TOX activities aim to fulfill six scientific and technology objectives:
Objective 1: Identify structural and physico-chemical characteristics that drive the toxicity of MeNPs.
The main objective is to perform a comprehensive survey of existing toxicity studies on two
MeNPs and construct a sound joined database (dataset) based on well-selected sources. Important consortia and toxicology labs actively involved in the field will be contacted to acquire access to their database(s). We will also make use of public sources, and available database from Health and Environmental Services. In these databases, relevant physico-chemical characteristics (PCCs) describing the material used and the measurable biological endpoints assessed will be combined. Based on the compiled “joined database”, classification models (Classifiers) will be generated based on key descriptors in order to classify the selected MeNPs into groups of similar toxicity patterns.
Objective 2: Identify the relation of MeNPs descriptors to toxicological endpoints and classify a training set of MeNPs into groups of regular toxicity patterns.
Based on this database, machine learning algorithms and multivariate analysis will be used to develop first dendrograms of hierarchical clustering of MeNPs and then a classifier model to define groups of ENPs having regular toxicological profiles. A successful classification will be achieved through identification of the different key descriptors (structural and PCCs) of MeNPs known to induce cellular toxicity. These descriptors will be suitably correlated to identify toxicological endpoints and used as readouts for building the computational models.
Objective 3: Development of a Computational Mechanistic Package (CMP) to predict MeNP toxicity.
A computational modeling package will be constructed to predict MeNPs toxicity. This consists in the implementation of an in silico model that integrates key cellular processes
underlying MeNP toxicity, including oxidative stress (ROS formation), mitochondria metabolism and bioenergetics, as well as various cellular cascades involved in cell permeability, DNA damage, and inflammation.
Objective 4: Develop a kinetic and dynamic model based on PBPK to predict MeNPs toxicity in human health.
In-vivo toxicity data from the database will be used to construct a kinetic PBPK model to
simulate the dynamics of ADME from different routes of MeNPs exposures. Such a PBPK model is significantly useful to evaluate kinetic of NPs mobility, disposition and associations between exposures and biomarker measurements in different physiologically relevant compartments (i.e. blood, urine, kidneys, liver, brain etc..).
Objective 5: Establish and utilize in-vitro & in-vivo innovative screening toxicity assays to calibrate and validate the CMP and the PBPK model.
We will design appropriate in vitro/in vivo assay models to generate experimental data needed to validate the computational models and perform a thorough characterization of the MeNPs used in the in vitro/in vivo assays. These data will be used to validate the computational models which will be developed.
Objective 6: Integrate all the testing strategies into a single modeling platform and demonstrate the feasibility of a MAP prototype.
The final step will be to integrate the constructed models into a platform, and to define an “overall estimate for toxicity / hazard potential”.
Project Results:
The MOD-ENP-TOX project developed a novel and generic modelling assay platform, which can be used as a risk indicator tool to predict the toxicity of nanoparticles (NP). This platform was based on in-depth analysis of two metal-based NPs, but which can be further developed to screen the toxicity of a large number NPs.
Below we will give a concise overview of the work done throughout the whole project. Firstly, an extensive database was constructed which includes both well characterized NP and their interaction with biological systems. The database was fed with data from literature, complemented with own experimental data. In parallel, different computational models were designed to unravel the cellular processes underlying NPs toxicity. These models included mechanistic, as well as kinetic and dynamic computational models. A new modelling platform was designed. Finally, the platform was tested and validated with the data extracted from the database.
1. Construction of a database and identification of key descriptors of the training set of metal-NPs and their classification into groups
Initially KU Leuven chose to populate the database with existing in vitro toxicological data
from the following sources: internal data of the S2Nano-project (the Belgian program “Science for sustainable development”); the ENPRA-project (a European project to develop and implement a novel integrated approach for engineered NP risk assessment); and literature data of peer reviewed papers. After investigation of the different data sources, it was decided to use data from published peer reviewed papers on the in vitro toxicity of amorphous silica nanoparticles and supplement these data with the internal toxicity data from silica nanoparticles of the he S2Nano-project. Until date, data of in vitro experiments on the most common investigated toxicological endpoints for NPs were collected; cell viability, apoptosis/necrosis, genotoxicity, oxidative stress and inflammation.
Different possibilities of how to store the curated data sets were reviewed. A human readable format was favoured to store and exchange data or datasets. Genedata introduced two human-readable database formats, the first GeneData Arrays (GDA) in order to store assay endpoints data and the second one GeneData Columns (GDC) to store the PCCs. These formats allow to be read by other software due to its logical structure. The GDA/GDC format can be used to import data into Genedata Analyst, software for data integration and analysis. Genedata Analyst will be used to perform tasks such as classification but also integration and grouping of data sets from different labs and different experimental setups. The consortium participated in the US Nano Working Group to discuss the new ISA-TAB-Nano format, which was published during the first year of the project. It was decided that the usage of GDA/GDC format supporting ISA-TAB-Nano nomenclature was the most efficient format to perform the upcoming data analysis steps and still being able to exchange data in ISA-TAB-Nano once a parser is written to convert the different file formats. The GDA/GDC file format was used to curate the data of publications which have been selected as the first approach to perform a toxicity classification of NP. After entering the data, the GDC/GDA files were reviewed and nomenclature and errors in these data could be easily identified using Genedata Analyst by checking for inconsistencies (e.g. spelling errors and duplicate data entries).
A detailed quality evaluation is needed both for the measured physico-chemical characteristics (PCCs) of the NP and their interaction with biological systems. As agreed with other consortia, ten PCCs that might determine the toxicity of NPs were included in the database: size distribution, agglomeration state, shape, crystal structure, chemical composition, surface area, surface chemistry, surface charge, solubility and porosity. Within the MOD-ENP-TOX project, these ten PCCs were divided in three groups of priority: obligatory (shape, size, composition and crystallinity), optional but very important (surface area, surface chemistry and surface potential) and optional (solubility, porosity and agglomeration/aggregation). In addition, also the reliability of the characterization method used was introduced into the GDC. A supplementary purity score and also an overall score for the amount of characteristics
specified or measured in the article were added. As there are no international standards for NP testing it is difficult to score the reliability of the assays used. Data evaluation is based on: the amount of information that is provided about the assay, the use of positive and negative controls and whether are not possible interaction of the assays with the NP were considered.
The development of a tool to classify the toxicity of NP requires a sufficient number of data points in order to define the toxicity of a NP based on the respective assay readout. A toxicity classification based on the PCCs needs a reasonable amount of NP which populate groups, toxic and a non-toxic. The cytotoxicity assays were chosen as a first group of assay readouts which were integrated as they contain different NP concentration levels which can be used as an indicator of level of toxicity. A viability drop of more than 25% was defined if a NP is regarded as toxic. The concentration level of the respective NP defines the level of toxicity. Therefore, four different toxicity levels were defined:
- Toxicity Level 0 (Not proven toxic (NPT)): a viability assay readout based on NP concentration >250 μg /ml annotates the NP as not proven toxic
- Toxicity Level 1 (Low Toxicity): a viability assay readout based on NP concentration >150-250 μg /ml annotates the NP as low toxic
- Toxicity Level 2 (Medium Toxicity): a viability assay readout based on NP concentration >75-150 μg /ml annotates the NP as medium toxic
- Toxicity Level 3 (High Toxicity): a viability assay readout based on NP concentration ≤75 μg /ml annotates the NP as highly toxic
The data subtracted from peer review papers as described above were generated in different labs without an underlying standard operating procedure. The heterogeneity of the data from the same viability assay readout is caused by several factors such as NP concentration units (e.g. mass per volume, mass per surface area, particle surface area per volume ...) cell type or contact time. Within the MOD-ENP-TOX consortium, a data integration strategy was setup on how to make these data sets comparable (if conversion factors are reported in the papers) and how to integrate the results. Although the importance of these features is not questioned, setting the focus on the main factors such as NP concentration was essential in order to perform a first toxicity grouping approach. After the available NPs have been assigned, a toxicity level based on their assay readouts, the impact of the PCCs of the NPs was analyzed. A Principal Components Analysis revealed that the PCC size is probably the best characteristic to differentiate between toxic and non-toxic NP. The individual PCCs were analyzed, but revealed no significant effect, except for the size. The effect of regrouping of the four toxicity levels (assigning a low toxic NP to the non-toxic group and thus reducing the number of different groups) was analyzed. Several classification terms were applied, the Decision Tree proved to be best suited for classification, which was verified by Genedata using a Leave-One-Out cross validation.
Data on the in vitro toxicity of amorphous silica NPs, particles with an amorphous silica shell and zinc oxide NP were collected from peer reviewed papers. Papers investigating particles with a defined shape (aspect ratio (AR) <3), composition (silica, silica shell or zinc), crystallinity (amorphous) and primary size were included. Apart from the obligatory characteristics (primary size, shape, composition and crystallinity), other physicochemical determinants of toxicity were also recorded: the surface charge, surface area, agglomeration/aggregation status, solubility and porosity of the particles. Five types of toxicological endpoints were included in the database: cell viability, apoptosis/necrosis, genotoxicity, oxidative stress and pro-inflammation. For data regarding in vivo toxicity of silica NP additionally included studies on the effect of the particles on mice or rat (experimental in vivo studies), and information on bio distribution, growth and consumption, blood cell count, energy metabolism, genotoxicity, pro-inflammation, oxidative stress or apoptosis/necrosis (see Figure 1) After extensive filtering based on the completeness of the data, only 5 papers were retrieved to be included in the database.
From the initial Medline search, which revealed 450 relevant papers, 70 papers were retrieved to construct the database, describing 147 different silica NP (or NP with a silica shell). For ZnO NP, a list of 36 papers was retrieved to construct the database, describing 70 different ZnO NPs (or NPs with a zinc oxide shell) and their characteristics in 229 different assays. Data were manually curated from text, tables and figures. To make the database exchangeable, data are stored in ISA-TAB-Nano, a standard format for sharing nanomaterial research data.
The database was further enriched by generating data from own experiments and feeding them into the database, in order to allow for validation of the computational models. Since the main toxicity pathway investigated in the database and the computational models to predict the toxicity of amorphous SiO2 and ZnO NPs is oxidative stress, both in vitro and in vivo experiments focused on the general toxicity and oxidative stress as a pathway underlying this general toxicity.
In vitro testing was done using the cell types which best illustrate the major ways NPs may enter and distribute in the human body such as epithelial cells and macrophages (inhalation), endothelial cells (blood circulation). First, the effects of different concentrations of silica NP were measured in viability assays (in human bronchial epithelial cells and human blood monocytes) via WST-1 reduction assays and the LDH release assays. The smaller the amorphous silica NPs the more cytotoxic they were for the epithelial cells and the monocytes. The cytotoxicity of the purpose made SiO2-10 and commercial available ZnO NP was tested on mouse lung epithelium cells. The MTT reduction assay was used to measure viability of the cells after 48 hours of exposure to different concentrations of the NPs. Again a dose-dependent response was observed after SiO2-10 and 10 nm ZnO NPs exposure.
To further investigate the mechanism responsible for the loss in viability of the cells, oxidative stress caused by the NPs was measured. Mouse lung epithelium cells were exposed to a sub-toxic concentration of ZnO NP for different incubation periods. After exposure reactive oxygen species (ROS) production was measured with the DCFH-DA assay. The intracellular ROS-production of the epithelium cells exposed to ZnO NPs for 6 and 16 hours was significantly higher than in cells exposed to culture medium. Since oxidative stress is closely related to inflammatory response by activating e.g. the nuclear factor-kappa B (Nf-κB) and p53 signalling pathway, the time- and dose-dependent effect of ZnO and SiO2-10 NPs on IKB-α and p53 proteins in these cells was investigated with western blot.
All these experiments allow quantifying the cellular response at different time-points. However, cell culture assays cannot reliable model the higher level interactions which are presented in living organism. Therefore, in vivo studies were needed.
Firstly, zebrafish embryo/larvae were exposed to SiO2-10 NPs, ZnO NPs, Zn(NO3)2 or vehicle in their medium. At different time point, they were washed and the amount of Si or Zn that they incorporated by the zebrafish was determined. No significant differences were observed for the different treatment groups. However, the results suggested an active uptake of the ZnO NPs and this in contrast with the Zn ions which are not taken up by the larvae. Also silica seems to be taken up by both the embryos and larvae.
Glutathione (GSH) is an important anti-oxidant of the cell, and is oxidized to GSSG when there is oxidative stress caused by reactive oxygen species. The ratio between GSSG and GSH is used as a measure for cellular toxicity. For the four different treatment groups the amount of GSSG and GSH per mg proteins was measured in the zebrafish 6 to 120 hours post fertilization. No significant differences were observed between the treatment groups.
Secondly, the bio-distribution of the NP in mice was investigated, as well as the effect on oxidative stress and inflammation pathways. Concentration of Zn and Si were measured at different time-points after the intravenous dosage of vehicle, 10 nm ZnO NPs, SiO2-10 NPs or Zn(NO3)2. The blood, brain, lung liver, kidney and spleen of the mice were analyzed with ICP-MS. No statistical differences were observed between different time-points and treatment groups after non-parametric tests. Reduced and oxidized levels of glutathione were also measured in these organs and at the same time-points. No statistically significant differences can be seen between different treatment groups due to the high variance within these groups. IKB-α and p53 concentrations were also measured in the liver, spleen, brain and the lung of the mice at the different time-points.
In a final stage, a full set of characteristics was reported for each NP used. The characterization is done in house to ensure that the same protocols were used for all NPs, and that the NPs have the desired PCCs analyzed. Since most techniques have some drawbacks, an array of techniques is needed to fully characterize the NP. The NPs were characterized with SAXS, ZP, DLS, TEM for size analysis. The PCCs determining the toxicity of NPs are: chemical composition and purity of the NPs measured by inductively coupled plasma mass spectrometry (ICP-MS), solubility measured by dialysis experiments, crystal structure measured by transmission electron microscopy (TEM), particle size distribution measured by TEM, small angle x-ray scattering (SAXS), dynamic light scattering (DLS), shape measured by TEM and SAXS and agglomeration and/or aggregation measured by TEM, SAXS and DLS. Surface charge was measured by zeta-potential measurement, surface chemistry, measured by Fourier transform infrared spectroscopy (FTIR) and surface loading, measured by thermogravimetric analysis (TGA).
Besides the measurements of the PCCs, their stability was tested in Danieau’s medium, since this medium was used to maintain zebrafish larvae which were used in the in vitro experiments as described above. This experiment ensured stable concentrations of the NPs throughout the duration of the in vitro experiments. Therefore, we monitored the NPs before and after they were exposed to the larvae and Danieau’s medium with both SAXS and DLS. The size monitored with DLS showed no difference in size before and after the exposure. Furthermore, the concentration was monitored with SAXS. These results showed that the concentration did not deviate after exposure.
Taken together, all these experimental data were added to the database and used to validate the computational models, as described below.
2. Development of a Computational Mechanistic Package (CMP) and a physiologically-based pharmacokinetic (PBPK) model to predict the toxicity of MeNPs in a quantitative way.
In parallel, three different computational models were designed to predict the toxicity of NP, which were fed with the data from the database described above.
Development of a computational model of mitochondria
A computational model of mitochondria was developed, which took into account appropriately balanced mass, charge, and free energy of various ions and electrons as well as the transduction systems. The model was based on a set of previously published data. Both the experimental data and computational models were previously developed for cardiac mitochondria.
The basic components we included in the new mitochondrial model are:
- The reactions of the mitochondrial respiratory chain for complexes I, III, and IV of the electron transport system and ATP synthesis for the F1F0 ATPase
- The substrate transporters, including adenine nucleotide translocase and the phosphate–hydrogen co-transporter
- Cation fluxes across the inner membrane, including fluxes through the K+/H+ antiporter and passive H+ and K+ permeation
- In addition, we integrated in our new mitochondrial model the production of the free radical superoxide (O2-), which is dependent on the NADH concentration and the mitochondria membrane potential (ΔΨm). Superoxide is transported from the matrix to the cytosol via a transporter called IMAC. The enzyme superoxide dismutase (SOD) catalyses the dismutation of superoxide (O2-) into hydrogen peroxide H2O2. H2O2 is highly toxic, decomposes with formation of free radicals, but it is rapidly neutralized and transformed into water by endogenous efficient defence molecules, e.g. the antioxidant glutathione (GSH) and the enzyme catalase (CAT).
It is important to notice that all the models published in the literature only considered one
component or a very small set of components of the mitochondria, and none previously integrated the different mitochondrial components together. By contrast, in our mitochondrial model, we integrated all the components needed for the respiratory chain and ATP production, the transporters, the mitochondrial permeability potential as well free radical production. In addition, our model is the only model, which integrates the effect of metal NP on the mitochondrial permeability potential (↓ΔΨm) and the oxidative stress chain (↓catalase).
Development of a computational model of p53 activation dynamics
Two deterministic models of the p53/Mdm2 circuitry were developed, based on previously published models reflecting biological reality with a minimum level of detail. We considered two different mechanisms for the DNA damage response:
- ARF activation, followed by sequestering of Mdm2
- ATM activation followed by phosphorylation of Mdm2 and p53
To simulate the effect of a gradual increase in DNA damage induced by MeNPs, we introduced a species called ROS (reactive oxygen species) into the model. The model predictions regarding the oscillation period are in agreement with experimental data. Under conditions of irradiation, this model generally produced one large initial peak of p53 after DNA damage, followed by irregular oscillations in which p53 levels remained relatively high.
Our model integrating the above biomarkers and their connections is unique. Indeed, ROS
input to the ARF and ATM was not published, and the in silico connection between mitochondria ROS production and p53 activation was not yet published. The dynamic of p53 oscillations remains unclear in experimental studies and oscillation is very difficult to measure experimentally. The in silico p53 model proposes a rational explanation for p53 oscillation dynamics: it depends on two stimulation modes dependent either on ATM or ARF. Such an in silico model may be useful for discovery of strategies (target and compound) aiming at modulating such oscillations or for supportive their use as a biomarker: p53 could become a new readout (biomarker) for MNP toxicity testing. Likewise, ARF and ATM models are generic models, which could be used for oncology research and for safety screening (effects of MNPs alone or in association with other molecules).
Development of a computational model of NFκB
Based on simplifying reductions of the IκB–NFκB signalling, a computational model was developed, describing the temporal control of NFκB activation by the coordinated degradation and synthesis of three IκB proteins isoforms (IκBα, IκBβ, and IκBε). The model demonstrated that IκBα is responsible for the strong negative feedback that allows for a fast turn-off of the NFκB response, whereas IκBβ and IκBε function to reduce the system's oscillatory potential and stabilize NFκB responses during longer stimulations.
NFκB is a protein complex that controls transcription of DNA. Based on simplifying reductions of the IκB–NFκB signalling, we developed a computational model that describes the temporal control of NFκB activation by the coordinated degradation and synthesis of IκB proteins isoforms (IκBα, IκBβ, and IκBε). The model demonstrates that IκBα is responsible for strong negative feedback that allows for a fast turn-off of the NFκB response, whereas IκBβ and IκBε function to reduce the system's oscillatory potential and stabilize NFκB responses during longer stimulations. Varied by two orders of magnitude on either side of the standard rate value of the three rates constant, IκBα transcription rate constant, IκBα nuclear import and rate constant for IKK degradation is a concrete example of how to use modelling to understand oscillation and the underlying mechanisms that delay them and could deregulate the pathway and induce toxicity and cell death.
Of importance, the results generated in silico, using the mitochondrial model, were validated against data of published experiments. To validate the model into a broader context and to become reliable to predict effect over time using different sizes of particles, the computational model needed additional test results of in vitro experiments. The limited (but crucial) set of tests were setup, including effects of MeNPs on diverse molecular targets implicated in cytotoxicity (or cyto-survival), such as inhibition of glutathione or blockade of complex I, in addition to downstream molecular targets, p53 and NFκB.
Development of a physiologically based pharmacokinetic model to predict MeNPs effects on human and demonstration of the feasibility of the model using subsets of MeNPs
A particle size-dependent PBPK model for ZnO NPs in vivo was developed. The model predicted dynamic bio-distribution of MeNPs. We estimated key physicochemical parameters of partition coefficient and metabolic/elimination rate based on our previously published data quantifying the biodistributions of 10 and 71 nm ZnO NPs and Zn(NO3)2 in different tissues of mice. Time-dependent partition coefficients were used and metabolic/elimination rates to determine the adequate PBPK model. The predictability of this model was assessed by calculating the mean absolute percentage error.
The PBPK models provides a potential effective tool to estimate the time course of chemical accumulation in target tissues of organism and can be incorporated into a quantitative risk assessment framework. Therefore, the development of a MeNPs PBPK model with the highly adequate predictive power is urgently needed to gain insights into underlying mechanical processes, hypothesis testing, and to guide experimental design. The advantages of a nanometal PBPK model not only can reduce animal testing and cost, but also can simulate and predict bio-distribution in human and human response to NP.
3. Validation and calibration of the computed models by quantitative measurements of toxicity in in vitro and in vivo assays and integration of all testing strategies into a modelling platform
Due to an unforeseen termination of participation of one of the consortium members, the original planning to validate and interconnect the three newly developed computational models needed to be revised. As a consequence, the core of the project was focused on one classifier model. The classifier model was validated on using experimental data in the database from 4 amorphous silica NP and its viability assays. As the advanced classification tool allowed the user to influence the toxicity group assignment of the NP used for the training of the classifiers by specific input parameters, the effect of the chosen parameters on the classification outcome was investigated. For the final validation of the classifier using experimental in vitro data from viability LDH release assay and WST-1 reduction assay the different input parameters were chosen to compare the results.
It was concluded that the new input parameters proved to have a high impact on the assignment of NP to the toxicity groups toxic and non-toxic. These input parameters were Viability Threshold, NP Toxicity Concentration Threshold, NP Non-Toxicity Concentration Threshold, Minimum Number of Confirming Measurements. The evaluation setup of comparing 26 runs with different settings confirmed the validity of the approach (proof of principle) and the expected impact of the parameters on the classification outcome. The Random Forest proved to be a suitable classification method for NP toxicity assignment. The classifier performance was shown to highly depend on the settings which determine the initial toxicity group assignment of the NPs and has been confirmed in most runs to deliver AUCs of ROCs of about 0.75. With optimal settings an area under the curve (AUC) of 0.87 was achieved, which indicates a fairly good prediction accuracy. A schematic overview is given in Figure 2.
Therefore, a number of improvements have been included in the new version of the model, since a small change in the viability threshold could have a strong effect on the population of toxic and non-toxic groups. The same applies to the NP concentration threshold at which a positive (toxic) assay output refers to a toxic annotation of the NP. To our best knowledge, there are no standard values when a viability assay is regarded to have a positive outcome. For that reason, the classification model has been extended to allow the user to create individual toxicity/non-toxicity NP groups depending on customized thresholds for both the viability and the NP concentration. Another feature has been added to the classification model to define the number of minimum data points necessary to label a NP as toxic or non-toxic. Although the data points from a variety of viability assay data from a multitude of publications are computed based on replicates and the potential of outliers should be excluded, it may still be possible that an error persists in the database. Most viability experiments are thoroughly set up with several NP concentrations in order to get more than just one measurement. However, increasing the minimum number of data points can significantly reduce the population of the toxic and non-toxic NP groups especially for values > 2.
The differentiation of four toxicity groups (described above, namely not proven toxic, low, medium and high toxicity) was based on the assumption that enough NP data is available to populate the groups for subsequent classification. As this assumption did not hold, we decided to differentiate only between two groups to provide more robust classifications using one concentration threshold 75 µg/ml.
The drawback of this setup is that such a threshold is crucial and very similar concentration levels can still result in two opposite assignments (e.g. 75 µg/ml and 76 µg/ml). Classification and cross-validation tests with regrouping of toxicity levels and excluding NPs assigned to a group of “insecurity”, however, seemed to perform better (data not shown). Toxicity groups NPT and Low Toxicity have been merged to non-toxic, High toxicity has been assigned as toxic. The medium toxicity group has been omitted as the group of “insecurity”.
Based on these results, we decided to develop thresholds for the two groups toxic and non-toxic and added the following assumptions:
Toxic NP Group: Viability < Viability Threshold (default: 75%) and NP Concentration < ConcTox Threshold (default: 75 µg/ml)
Non-Toxicity NP Group: Viability ≥ Viability Threshold (default: 75%) and NP Concentration > ConcNonTox Threshold (default: 75 µg/ml)
These two threshold parameters ConcTox and ConcNonTox are exposed to the user and will have a high impact on the toxicity group assignment. Due to the added functionality which allows for dynamically changing the settings described above, the toxicity assignment has to be performed for every single classification run. Therefore, the extracted data set without prior NP toxicity assignments needs to be available.
Subsequently, the new functionality was evaluated by running two classification evaluations with the extracted viability data set of the database for silica NP with the varying preset parameters of viability threshold, NP toxicity concentration threshold, NP non-toxicity concentration threshold and the minimum number of data points. Both evaluation models confirm the good performance of the Random Forest classification method. While the settings used in Evaluation A (viability threshold of 50% and a NP toxicity concentration threshold of 75 µg/ml) provides a better estimate of toxicity for the underlying data set, Evaluation B (viability threshold of 75% and a NP toxicity concentration threshold of 150 µg/ml) proves that a different choice of parameter settings does not have the same impact on the classification models.
As the Random Forest performed best on the underlying data set, this classifier was used in order to predict toxicity for all NPs to be analyzed. The result of the cross-validation has been added to the toxicity prediction. As the user did not have influence on the underlying data set for the validation, the result image was static and did not change.
Due to our changes and added functionality explained above, the underlying data set is now dynamically created depending on the settings of the user. Also the next step, i.e. the evaluation of the classification performance is now performed dynamically based on the user-defined validation dataset. In each run of the tool the available classification methods are compared on the validation dataset by means of nested cross-validation. The best performing prediction model is then used for classifying the NP under investigation based on the respective PCC measurements. The outcome of the comparative performance evaluation and the classification of the input NP are presented in an HTML report. This report includes a dynamically created image of the averaged ROC curves, which were used for assessing the prediction accuracy of the constructed models, as well as a table specifying the estimated probability that the input NP is toxic and non-toxic, respectively.
Taken together, the advanced classifier proved to cope with different experimental setups and allows to assign toxicity for different thresholds on viability as well as NP concentration thresholds. Although the validation of the classifier was limited to four NP, 2 different assay readouts and 6 different NP concentration steps, the classifier proves to provide comparable toxicity estimates. Both the classifier estimate output and the experimental data give proof that the change of thresholds used to define NP toxicity has a huge impact on the outcome. The classifier model returned an estimate of the toxicity of a NP under investigation based on its PCC’s. The model was trained on data from 116 NP measured in different viability assays.
4. Modeling platform infrastructure
The classifier model is accessible via a web user interface. On the main page a screen is offered to the user at which five parameters (five common particle characteristics) need to be filled out: Diameter – or size (in nm), Specific Surface Area: (m²/g), External Surface Area: (m²/g), Zeta potential in water (mV) and Hydrodynamic diameter in water (nm).
After clicking the button ‘Run script’ an additional tab is opened showing the outcome of this specific simulation. Two different types of outcome data are provided: The statistical probability calculation whether the material (of the simulation) will be toxic (meaning inducing cytotoxicity (reduction of more than 25 %) in vitro after 24 h at a concentration of 75 µg/ml or less). This probability is given in percentage. The performance of the evaluated classifier algorithms is compared using a ROC (receiver operating characteristic) curve. Also the algorithm used to generate the given probability is specified.
Based on the validation data set of toxic and non-toxic NPs, which has been extracted from the database according to user-defined toxicity parameters, the classification performance of five state-of-the-art supervised classification algorithms (Random Forest, K-Nearest Neighbors, Nearest Shrunken Centroids, Weighted Voting and Adaptive Boosting) is determined in a nested cross-validation. The best classifier is selected based in the highest average area under the ROC curve. This classifier is then applied to predict the toxicity of the NP under investigation. The returned prediction score is transformed into a value between 0 and 1 which corresponds to the probability that the input NP is toxic. The outcome of the method comparison and the classification of the input NP are presented as shown in Figure 3.
The modelling platform setup allows access management using a name and password. Results from mechanistic modelling runs were stored and could be reopened using results management infrastructure. Users can register or log in if they already have credentials to run a classification model. In addition, the new platform also allows to store results from the classification model as well as the input parameters. The generated results can easily be downloaded or shared with other users. The info button gives also access to valuable information including when a modelling job has been started and when it has been finished.
Potential Impact:
The MOD-ENP-TOX project aimed to develop a novel and generic modeling assay platform, which can be used as a risk indicator tool to predict the toxicity of metal-based NPs. This platform was based on in-depth analysis of two MeNPs, but which can be further developed to screen the toxicity of a large number ENPs. To illustrate the impact of the MOD-ENP-TOX project on European research and industry, we will describe how the project contributed to the enhanced the current scientific knowledge, but also how the project led to new collaborations and future research projects.
Firstly, an extensive database was constructed which includes both well characterized NP and their interaction with biological systems. The database was populated with experimental data from peer-reviewed papers on the in vitro toxicity of spherical amorphous silica NPs, spherical NPs with an amorphous silica shell and zinc oxide NP. The database is set up following the internationally agreed format ISA-TAB-Nano. This will allow external groups to extract and use data from the database format. Due to this standardized format and high quality (including criteria were used to guarantee the quality of the data in the database), the database is not only useful within the MOD-ENP-TOX project but can also be used by other stakeholders who wish to determine quantitative structure/activity relations for nanomaterials. Therefore, the full database will become public on our server after 1-2 years after our internal exploitation.
In a second stage computational mechanistic models were used to unravel the cellular processes underlying NPs toxicity. The main challenge in predicting toxicity using classifiers based on literature data comes from the fact that assays and readouts used for toxicity assessment as well as determination of PCCs are diverse, and comparable only to a very limited extend. A modelling approach, as presented within MOD-ENP-TOX, has to cope with a sparse data matrix (i.e. the exact same readout is available only for very few nanoparticles) which makes assumptions necessary under which circumstances a result shall be regarded “positive” or “negative”, respectively. The framework presented here operates with a set of pre-defined rules (e.g. a nanoparticle classified as “toxic” in one assay will always be regarded as toxic, even if another assay classified it as “non-toxic”) and user-defined thresholds, that allow for different levels of stringency, safety margins, and required supporting evidence for categorization. The framework can be used to test the performance of classification under different assumptions and, more importantly, can be easily adapted to a growing and hopefully more and more standardized data basis of nanoparticle properties and toxicities.
The consortium members actively participated different European initiatives throughout the project, contributing to the enhancement of the knowledge about risk assessment for the use of nanomaterials. Nanotechnology produces engineered nanomaterials having new or enhanced physico-chemical properties in comparison to their micron-sized counterparts. Some of these properties, like the high surface area to volume ratio, make them potentially dangerous to humans. To promote the development of a new generation of nanomaterials that are safe-by-design, an understanding of the relationship between the structure of the nanomaterial and biological activity is needed. In this context the Quantitative Nanostructure-Toxicity Relationships computational modelling technique is an effective alternative to experimental testing since it enables the prediction of (eco)-toxicological effects based on nanomaterial structure only.
The construction of these models requires the integration of expertise of nanomaterial scientists, (eco)-toxicologists, and modellers from academia, regulatory agencies and industry. Therefore, a network for trans-disciplinary cooperation was needed, which resulted in the COST action entitled ‘Modelling Nanomaterial Toxicity (MODENA).
This COST action was established to promote and to co-ordinate these inter-disciplinary collaborations, with the ultimate aim of producing computational models for engineered nanomaterials. Peter Hoet of the KU Leuven is one of the responsible persons representing Belgium in these activities.
As a scientific outcome of these actions, a new collaboration has been setup between Sabine van Miert (Thomas Moore, Geel Belgium), Peter Hoet and Jean-Pierre Locquet (KU Leuven, Belgium). The database developed by the MOD-ENP-TOX project will be used by the modeling groups of Sabine Van Miert to exploit in more detail PBPK models and classifier models. These efforts will be done outside the contact of the project but will give the data collected in the project an ‘new’ live since the potential of the whole database was not yet fully exploited. Secondly, Sabine van Miert (Thomas Moore, Geel Belgium), Peter Hoet (KU Leuven, Belgium) and Helena Oliveira (University Aveira, Portugal) are building a PBPK model for AgNPs (no MOD-ENP-TOX data).
List of Websites:
http://fys.kuleuven.be/apps/modenptox/
