Skip to main content

Predictive toxicology of engineered nanoparticles

Final Report Summary - PRENANOTOX (Predictive toxicology of engineered nanoparticles)

Executive Summary:
Nanotechnology is an emerging field that offers among others a fast and targeted design of material properties. As new types of nanomaterials (NM) find application in different areas the need to understand their potential risks becomes urgent. However, experimental toxicological testing of nanoparticles (NPs) is time-consuming and costly. Thus, the turn towards better basic understanding the relationship between the physicochemical properties of NM and their potential toxicity, modeling the observed relationship and being able to predict it have become a focus of the research activity in the field of nanosafety. However, the modeling process requires the availability of large databases which are currently absent. Thus the approach chosen by PreNanoTox project to address the above goals relied on the formation of three synergistic lines of activity:
(1) Formation of a large database on the relationship between physicochemical properties of NM and their biological end points based on automatic information extraction (IE) of data from peer refereed articles
(2) Understanding some basic rules on the relationship between the size of nanoparticles (NPs) and the ability of the cell membrane to wrap them NP upon interaction.
(3) Develop models on the relationship between the physic-chemical characteristics of NPs and their potential toxicity.

Thus, our first goal was to develop tools for information extraction that will enable the creation of a unified database from peer-refereed scientific articles, providing the basis for improved modelling and predictive nanotoxicology. The major challenge was to develop tools for extracting textual and graphical information from peered refereed articles, with the ultimate aim of forming a big data base in the domain of nanotoxicology. The textual mining of literature has been addressed in other domains, but not in the nanosafety one. Moreover, to the best of our knowledge the challenge of figures analysis is rarely addressed in the literature, and seems to be so far not fully resolved, as it requires the merging of different approaches including image analysis, optical character recognition (OCR), text mining, and data integration to accomplish it. For text mining we employed a state of the art tool, the Stanford Parser to generate output for every sentence: a syntactic parse tree and the dependencies between words and phrases. The dependencies enabled us to connect an entity to its attribute, and to extract correct facts. The main entity types were: nanoparticle, cell model, animal, measurement methods, and assays. To estimate quantitatively the performance we use the following measures: (1) Precision (the ability to filter out; the probability that the selected information is relevant). Current result: 85%; (2) Recall (the ability to generalize; the probability that the relevant information is specified). Current result: 37%; (3) F-measure (estimation of precision and recall). Current result: 51%. It should be stressed that the goal was to obtain an F-measure of 60% for NPs. Specifically, the accuracy accepted targets for text information extraction are: (1) Extract 60% of the entities, with 85% precision and extract 30% of the attributes, with 75% precision. We have addressed automatic figure analysis by laying down basic elements required for the analysis of the bar-chart, the most generalized type of graph, but did not reach the state of full validation of the developed algorithm. Since there are no benchmarks or evaluation frameworks in the field of figure processing it is impossible to present a definitive success level (a strictly defined success level is acknowledged in the area of recommender systems, where Netflix contest made it possible to conduct benchmark tests on the same datasets). The successful detection is demonstrated in elaborated tables that compare the output with the original values.

The second goal of PreNanoTox activity addressed better understanding of the underlying mechanisms of the primary interaction of nanoparticles in the cell membrane. The interaction of nanoparticles with membranes is a necessary first step for any toxic effects of nanoparticles in living cells. While small nanoparticles may penetrate a lipid membrane, larger nanoparticles with diameters above 20nm either bind and attach to or even get wrapped by the membrane. Our suggestion was to apply appropriate theory and simulation assuming that the surface chemistry of a nanoparticle and its size determine the strength of the non–specific adsorption of a nanoparticle to a cell surface, leading beyond a certain adhesion strength threshold, to efficient uptake of the nanoparticles. We have addressed this challenge by describing wrapping of nanoparticles using a continuum membrane model, where the curvature-elastic properties of the membrane are taken into account by a set of curvature-elastic constants. Our main objectives were a systematic characterization of nanoparticle adsorption to and wrapping by membranes, and the investigation of the dynamics of wrapping to obtain information about pathways and typical times scales. We present our results for the systematic characterization in form of phase diagrams that show the wrapping state of the nanoparticle for various values of the curvature-elastic parameters of the membrane, for various nanoparticle characteristics (e.g. shape, local curvature, aspect ratio), and for different nanoparticle-membrane adhesion mechanisms. To achieve these goals, we have employed analytical theory, numerical calculations using triangulated membranes, and experiments on the interaction of nanoparticles with red blood cells and with lung cells to validate the predictions obtained by theoretical modeling and simulation.
For nanoparticles that get wrapped by homogeneous lipid-bilayer membranes, the deformation energy of membranes can be characterized by their bending rigidities and tensions and—for homogeneous membranes—the nanoparticle-membrane interaction is characterized by the adhesion strength. For various particle shapes and membrane elastic properties, we have calculated phase diagrams that delineate regions in parameter space where the particle is unbound, partially attached, or completely wrapped. We show that for non-spherical nanoparticles also shape and local curvature, as well as nanoparticle orientation with respect to the membrane are important. For multi-component membranes the wrapping behavior is more complex. If the components form domains, for a domain that is larger than the nanoparticle surface area the line tension at the domain boundary assists wrapping, while for a domain that is smaller than the nanoparticle surface area the domain boundary can lead to frustrated endocytosis of nanoparticles. If the components of the membranes do not phase segregate, the entropy of the molecules in the membrane that bind to the nanoparticle (for example receptors that bind to ligands anchored to the particle) has to be taken into account. We find stable partial-wrapped states for high receptor-ligand bond energies and low receptor densities. Similar, partial-wrapped states are also obtained for charged membranes that interact with charged nanoparticles in a biological environment where Coulomb interactions are usually strongly screened. Interestingly, we find that charged nanoparticles are not only wrapped by oppositely charged membranes, but also by membranes that contain charged molecules and are overall neutral or even slightly like-charged. A cytoskeleton mechanically stabilizes cells and can also actively contributed to nanoparticle uptake, either if the nanoparticles aggregate and if the aggregates are large enough to be phagocytosed or possible also for on smaller scales as hinted by our work on cells that adhere to nanostructured surfaces. A passive cytoskeletal network that stabilizes the lipid bilayer and that provides a direct steric barrier for the nanoparticles stabilizes partially-wrapped states for particles that are less than half wrapped and suppresses states for particles that are more than half wrapped.
Experimental studies with human red blood cells and lung cells to validate the theoretical results show that nanoparticles adhere to only 2% of the cell area and bind irreversibly and electron micrographs show partial-wrapped nanoparticles attached to the cells. This corresponds to the regime for small densities of adhesive sites and a strong binding energy described above for that we also theoretically expect partial-wrapped nanoparticles. Partially wrapped nanoparticles that attach to the outside of the cell act as cup formers, theoretical predictions show that a shape change of the red blood cells from discocyte to echinocyte is possible, in agreement with the experimental observations.

The third goal of our activity was to extend the traditional QSAR paradigm to the field of nanotoxicology. This was made possible by linking appropriate descriptors with emphasis on those which determine the strength of adsorption of nanoparticles to cells, with biological responses. The challenge was to develop a modeling tool, to identify the threshold of concern for the toxicity effects of the nanomaterials, clustering of nanoparticles and to identify the eventual relationships between the activity and the properties of the nanomaterials. In order to reason about the use of the candidate descriptors for the QNAR models, we first evaluated the content-specific nature of the descriptors more suitable for the case of silica and then based on a chemometric approach, examined the possible descriptors, for which more data than those available for the silica were needed. Some key parameters playing a role on the effects of nanomaterials have been identified, in addition, an extended number of papers have been evaluated to verify the initial results based on the restricted set of evidence. To identify a suitable strategy for nano-descriptors, to be applied for material of interest, eclectic data are found to be more suitable compared to the traditional physicochemical descriptors. In particular, chemical composition and structure, in addition to size, shape, etc. appear to be potentially useful to characterize and model nanoparticles of different nature. CORAL has been advanced as a flexible tool, suitable to organize descriptors of eclectic nature within the same algorithm. It was possible to generate descriptors useful to build up models for nanomaterials. The identification of relevant ontologies through clustering of nanoparticles has become possible trough out a number of classification and clustering tools available on the internet. The most effective tools for this purpose have been identified within the QNAR and data mining modeling work package (WEKA, CARROT2, APACHE MAHOUT, etc.). To provide homogeneous data for the following activities the collected publications have been classified on the basis of cell type (e.g. macrophages, fibroblasts, ect.) and cell line assay (e.g. MTT, WST-8, etc.). The gathered data on silica nanoparticles and their relative toxicity and uptake mechanisms have been evaluated in order to find relevant features associated to cytotoxicity. The main physico-chemical properties are atomic decomposition, polarity, size and zeta potential. The optimal descriptors which are a tool of the traditional QSPR/QSAR analyses have been suggested by us in order to build up QSPRs/QSARs models by the means of CORAL. In this way some new models for toxicity of nanomaterials have been developed and some important features/parameters/characteristics of NPs associated to toxicity are identified. With regard to QNAR and data-mining modeling, the Monte Carlo technique has been used to build up two predictive models based on QSARs for the prediction of:(i) cell membrane damage of metal oxide nanoparticles; (ii) cellular viability (CV%) towards silica nanoparticles. Principally, the system of building up prediction for endpoint related to nanomaterials based on quasi-SMILES is suggested and examined for different endpoints related to biological effects.

Project Context and Objectives:
The fast development of nanotechnology provides us with new types of nanomaterials which find applications in many different areas of our everyday lives. While nanotechnology offers a fast and targeted design of material properties, experimental toxicological testing addressing the potential health risk associated with exposure to nanoparticles (NPs) is rather time-consuming and costly. This requires addressing several critical bottlenecks in developing and applying new computational modelling which will enable to predict and attain the goal of predictive nanotoxicology.
In order to develop a platform for predictive nanotoxicology, PreNanoTox addressed the following objectives:
(1) The current lack of centralized database severely limits the capability of the informatics experts to envision and analyze the relationships between physico-chemical properties of NPs and their different effects when interacting with living organisms and the toxicological endpoints induced by them. To address this need, we suggest developing approaches of automatic information extraction in the domain of nanosafety, which will enable to form a very large and unified database from peer-refereed scientific articles, and provide the basis for the following step. To the best of our knowledge this is first attempt of its kind to construct a unified data-base from scientific literature in the field of nanosafety.
(2) Extension of the traditional Quantitative Structure-Activity Relationship (QSAR) paradigm from chemicals to NPs by identifying and calculating appropriate NPs’ descriptors and by building externally validated Quantitative Nanostructure-Activity Relationship (QNAR) models linking NPs descriptors and observed biological responses.
(3) The ability to proper QNAR and data-mining modeling requires a mechanistic understanding on the relationship between the physico-chemical properties of NPs and their respective toxicity. Currently there are very few models of nanoparticle toxicity, especially those relating to dissolution of metal and metal-oxide NPs due to the inverse relationship between their surface activity and their size. For NPs, which do not undergo dissolution we suggest a mechanistic model where NPs' toxicity depends on the strength of adsorption of a NP to a cell surface which is determined both by NP's size and its surface chemistry. We intend to lay down the theory and simulate the adsorption and wrapping the NPs by the cell membrane, thereby providing a mechanistic basis for NP's toxicity, for the successful application of QNAR and data-mining modelling.

Project Results:
• The quantitative performance of Text mining resulted 85% precision, 37% recall yielding overall performance (F-measure) of 51%.
• We have addressed automatic figure analysis by laying down basic elements required for the analysis of the bar-chart, the most generalized type of graph, but did not reach the state of quantitative validation of the developed algorithm.
• Individuation of parameters that influence the cytotoxicity of nanoparticles. Initial work on a pilot number of papers, to evaluate some key parameters playing a role on the effects of nanomaterials.
• Evaluation of an extended number of papers, large enough for clustering, aiming to verify and check the initial results based on the restricted set of evidence.
• Definition of thresholds for some types of cells. The cell lines under investigation are lung cells A549, macrophages RAW 247.6 fibroblasts. For example, for HeLa cells exposed to pure silica NPs size of 45 nm has been identified as threshold for cytotoxicity. (Silica Nps < 45 nm are cytotoxic, while in the range of 45 and 500 nm they are non cytotoxic.)
• Determination of correlation rules between physico-chemical characteristics of nanoparticles and cytotoxicity. For example, the positively charged NPs are more cytotoxic than those negatively charged, except in the case of macrophages.
• Adsorption isotherms of model nanoparticles (polystyrene nanoparticles of 27, 45 and 100nm with carboxylate modified surface) to red blood cells (RBCs) and lung cells can be fitted by Langmuir model. Parameters of NP binding, such as binding constant and maximal extent of adsorption can be calculated from the mode.
• If one assumes similar binding sites for NPs on RBCs and lung cells, the density (number of binding sites per unit area) of binding sites on lung cells is ~13 fold larger than on RBCs, though the ratio of cell membrane areas is 25 (Alung/Arbc ≈ 25). This suggests more binding sites for NPs on A549 lung cells as compared with RBCs.
• The small amount of bound nanoparticles to RBCs, the measured dependence of the number of bound nanoparticles on the nanoparticle concentration in bulk, and the equal area of the membrane that is bound to nanoparticles with different sizes is not compatible with the theoretical models that use a homogeneous adhesion strength between nanoparticles and membranes.
• Using a kinetic model for the receptors, we extract a receptor diffusion coefficient from the dependence of the number of adsorbed nanoparticles on the nanoparticle concentration in bulk. A receptor model furthermore predicts for high receptor-ligand bond energies and small receptor densities partial-wrapped states for spherical nanoparticles as observed in the experiments.
• A receptor model predicts for high receptor-ligand bond energies and small receptor densities partial-wrapped states for spherical nanoparticles as observed in the experiments. The small receptor density required for partial wrapping is consistent with the experimental finding that particles adhere to only 2% of the total membrane area at the highest nanoparticle concentrations.
• Out of the three sizes the smallest NPs (27nm) were the most toxic one in causing shape transformation and deformability of RBCs at high NPs concentration. Shape changes are induced by NPs via changing spontaneous membrane curvature upon adsorption. Our calculations predict that large amounts of small nanoparticles adsorbed at the outside of the cell lead to shape changes from discocytes to stomatocytes, which is confirmed by the experimental observations.
• Though NPs are dispersed as single particles in the external medium, they tend to aggregate on the cell membrane. Longer of NPs adsorbed to the membrane leads to the formation of higher sizes of aggregation. This predicts that the longer a NP is attached to the cell surface the higher is its tendency to be incorporated in a large aggregate, with smaller efficacy to undergo uptake.

Potential Impact:
• 19 peer refereed article were published by the PreNanoTox consortium
• The Partners of PreNanoTox delivered 45 presentations in differents Conferences and meetings.
• New cell line oriented data base of nanomaterials with their cytotoxicity properties
• New models for toxicity of nanomaterials.
• Important features/parameters/characteristics of NPs associated to toxicity

List of Websites:

Coordinator: Prof. Rafi Korenstein,
Dept. of Physiology and Pharmacology,
Faculty of Medicine,
Tel-Aviv University, Tel-Aviv, Israel