A Strategy Towards the Generation of Testable Adverse Outcome Pathways for Nanomaterials

Manufactured nanomaterials (NMs) are increasingly used in a wide range of industrial applications leading to a constant increase in the market size of nano-enabled products. The increased production and use of NMs are raising concerns among different stakeholder groups with regard to their effects on human and environmental health. Currently, nanosafety hazard assessment is still widely performed using in vivo (animal) models, however the development of robust and regulatory relevant strategies is required to prioritize and/or reduce animal testing. An adverse outcome pathway (AOP) is a structured representation of biological events that start from a molecular initiating event (MIE) leading to an adverse outcome (AO) through a series of key events (KEs). The AOP framework offers great advancement to risk assessment and regulatory safety assessments. While AOPs for chemicals have been more frequently reported, the AOP collection for NMs is limited. By using existing AOPs, we aimed to generate simple and testable strategies to predict if a given NM has the potential to induce a MIE leading to an AO through a series of KEs. Firstly, we identified potential MIEs or initial KEs reported for NMs in the literature. Then, we searched the identified MIE or initial KEs as keywords in the AOP-Wiki to find associated AOPs. Finally, using two case studies, we demonstrate how in vitro strategies can be used to test the identified

started by the OECD Working Party on Manufactured Nanomaterials (WPMN) to support the development of future AOPs for NM RA and categorization. The AOP framework offers great advancement to RA and human hazard assessments of NMs, which often diverge from classical dose-response relationships and exhibit particulate-specific toxicity. Indeed, even small changes in their physico-chemical properties may significantly impair the nano-bio interface, aggravating predictability of traditional RA tools and methods.
It has been discussed that existing AOPs for chemicals can potentially be used for NMs as these may share similar KEs with chemicals (Ede et al., 2020). The AOP provides the mechanistic representation of an AO initiated by a MIE, thus reflecting the molecular level, and making possible the connection with the NM's physico-chemical properties via in silico tools, such as quantitative analyses of structure-activity relationships (QSARs).
While extensive efforts have been made towards the development of AOPs for chemicals, AOPs specific to NMs are still scarce. At the end of 2020, the OECD WPMN, as part of the Na-noAOP project, reported a methodology to identify, analyze and evaluate existing scientific data to prioritize NM-relevant KEs and so to contribute to the development of a knowledge base to inform AOP development and assessment for NMs (OECD, 2020). One of the main outcomes of this project was the development of a case study on the inflammation pathway and analysis of a specific KE for this pathway to establish an approach to advance future NM-related AOPs (OECD, 2020). Gerloff et al. (2017) in his study attempted to merge existing chemically induced liver fibrosis AOPs and proposed a putative AOP for metal oxide NMs by combining in vivo and in vitro literature data obtained for TiO 2 and SiO 2 NMs. However, the potential application of chemical AOPs to NMs is not comprehensively explored. prone to biased results (Ostermann et al., 2020). Moreover, even small variabilities in the physico-chemical properties of NMs have been shown to influence the toxicological outcome, which further challenges the grouping and read-across analysis of NMs. The development of intelligent and more efficient methodologies with lower costs is therefore urgently needed for hazard and risk assessment (RA) of NMs.
The adverse outcome pathway (AOP) framework can significantly support the advancement of RA approaches by developing predictive methods that utilize mechanistic and evidence-based data. The AOPs, first described a decade ago (Ankley et al., 2010), refer to conceptual structures portraying biological failures initiated by the interaction of a chemical with a biomolecule or biosystem that can perturb normal biology, impairing critical function and leading to adverse outcome(s) (AO) at organism or population level.
As shown in Figure 1, AOP comprise a series of key events (KEs) along a biological pathway from the molecular initiating event (MIE) to the AO. As such, the AOP framework provides systematic knowledge about key toxic mechanisms, thus being very effective at characterizing the individual biological impact and toxicological potential of substances and significantly improving the prediction of adverse effects.
Worldwide, there are many initiatives for further development and advancement of the AOP framework; at European level, the OECD has made significant efforts in this direction and initiated an AOP Development Programme in 2012. In collaboration with the U.S. EPA and the U.S. Army Engineer Research and Development Center, the EC's Joint Research Centre launched in 2014 the AOP Knowledge Base (AOP-KB) as a web-based tool encompassing the e.AOP.Portal, the AOP-Wiki, Effectopedia, the AOP Xplorer, and the Intermediate Effects Database (Delrue et al., 2016). In 2016, the NanoAOP project (OECD, 2020) was An AOP is triggered by a molecular initiating event (MIE), an initial interaction with a biological target (Anchor 1), which leads to a sequential cascade of cellular, tissue and organ responses (key events), linked to each other by key-event relationships (KER) to result in an adverse outcome (AO) of regulatory relevance.
Thus, we aimed to generate simple and testable strategies for the development of AOPs for NMs that are of relevance for human health. Our approach is based on using existing AOPs to predict if a given NM has the potential to induce a MIE leading to an AO through a series of KEs. Firstly, we identified potential MIEs reported for NMs in the literature. Then, we searched in the AOP-Wiki using the identified MIEs (as keywords) to find associated AOPs that can be applied to NMs and can be verified using in vitro and in silico approaches through testing the involved KEs.

Methodology
The first step towards generating testable AOPs for NMs was a literature search using the following scientific databases: PubMed, Embase, Scopus and Web of Science. The search, performed in the period until 1/12/2020 using the key words "adverse outcome pathway" OR "AOP" AND "nano*", resulted in 960 papers in total. After careful analysis and refinement on duplicates, reviews, AOPs/AOs reports, and type of organisms studied therein, 32 papers that covered both in vitro and in vivo studies on mammals were selected for further analysis. Next, each of the selected papers was evaluated by the software-based tool ToxRTool (European Commission, 2013), which assesses the reliability of in vivo or in vitro human toxicity studies. According to the criteria described by (Klimisch et al., 1997), this evaluation revealed that only 15% of selected papers provided reliable data with some restrictions, i.e., data are potentially useful but their relevance should be checked for intended purpose. The rest of the papers (85%) were evaluated as providing reliable data without restrictions. The papers then underwent data extraction, which included the identification of NM-induced AOs and MIEs relevant for NMs. In this study, we focused on the identification of initiating events relevant for NMs because chemicals may share common KEs with NMs, but major differences lie in their way of interaction with biological targets. To achieve this, the AOs and the respective first event (mainly molecular/cellular level key event) re-ported in each study were summarized. When analyzing these data, we found that certain events were consistently reported. Finally, the identified initiating events were used to search the AOP-Wiki for potential AOPs applicable for NMs, and all the AOPs linked to each of the used keywords were summarized. Since inhalation and ingestion are the primary routes of NM exposure, we focused on lung and liver fibrosis to describe our strategy to generate testable AOPs.

Identification of (molecular) initiating events relevant for NMs
To identify potential MIEs relevant for NMs, AOs reported in each of the selected research papers and their respective reported/identified first event were consolidated as presented in Table  1. The KEs can be described as a measurable change in the biological state representing an essential event for further biological effect(s) and progression towards the AO, but not bridging levels of biological organization. The critical step in AOP development is the identification of a MIE that is defined as "the initial interaction between a molecule and a biomolecule or biosystem that can be causally linked to an outcome via a pathway" (Villeneuve et al., 2014). In this definition, "a molecule" can be replaced by a NM, but the chemistry of the MIE should be carefully described to provide a coherent link between the physico-chemical properties of NMs and MIEs that is stronger than the links to adverse endpoints (Allen et al., 2014). When analyzing the extracted data from papers (Tab. 1), five potential MIEs for NMs were identified: (i) Interaction of particles/fibers with cell membranes/ biomolecules, (ii) reactive oxygen species (ROS) formation/ generation, (iii) lysosomal injury/damage/disruption, (iv) DNA damage/methylation, and (v) inflammation induction. All these initial KEs were obtained from both in vitro and in vivo studies. Instead of MIE, the term "initial key event (initial KE)" is used in subsequent sections because not all identified events occur at the molecular level.

Case study 1: Lung fibrosis
AOP 173 ( Fig. 3; substance interaction with lung epithelial and macrophage cell membrane leading to lung fibrosis) has been the most discussed AOP for its potential application to NMs. Briefly, the interaction between the substance and components of the cellular membrane (MIE) leads to the release of pro-inflammatory mediators (KE1) that promote the recruitment of pro-inflammatory cells into the lungs (KE2). Persistent inflammation leads to the loss of alveolar capillary membrane integrity (KE3) and activation of the adaptive immune response (T helper type 2 activation) (KE4), during which anti-inflammatory and pro-repair/fibrotic molecules are secreted. The repair and healing process stimulates fibroblast proliferation and myofibroblast differentiation (KE5), leading to synthesis and deposition of an extracellular matrix or collagen (KE6), and eventually lung fibrosis (AO). It appeared that some of the components of this AOP cannot be replaced with in vitro cellular assays (such as KE2, KE3 and KE4). NMs, particularly carbon nanotubes (CNTs), were shown to induce lung fibrosis in vivo via different interactions and pathways. When mining the literature, we found a recent comprehensive pathway analysis of in vitro results relating to multi-walled (MW) CNT-induced lung fibrosis (Vietti et al., 2016). Based on this information, we propose an AOP consisting of the major KEs that can be tested/verified under in vitro settings. Figure 4 shows the aligned initial KE-KEs-AO pattern that can be measured in vitro to predict the lung fibrotic responses in vivo and different strategies to test the potential of a given NM to induce an AO re-

Identification of potential AOPs in the AOP-Wiki
The AOP-Wiki "Key Events" module of the freely accessible web-based AOP-KB was used to identify AOPs applicable for NMs. The search revealed several titles linked to initial KEs identified in the first step (Tab. 1), e.g., the results for "Interaction of particles with cell membranes" as depicted in Figure 2. Then, the AOPs linked to each of these titles were retrieved: Table 2 shows the potential AOPs found in the AOP-Wiki linked to each of these initial KEs. We did not use the term "inflammation" in the search as it was widely recognized as a KE rather than an initiating event (Halappanavar et al., 2019). A detailed analysis of titles linked to different keyword searches and identification of associated AOPs is provided in File S1 1 .

Generation of testable strategies using simple in vitro/in silico experiments
In this work, initial KEs are considered as one critical component that can be shared by more than one pathway. The sequence of intermediate KEs connecting the initial KE with AOs should be described, including the definition of the biological state, methods used for intermediate KE observation and measurement, as well as evaluation of taxonomic applicability of a particular KE (Villeneuve et al., 2014). Another important AOP component is the KE relationship (KER) that is supported by empirical evidence and establishes directed and quantitative relationships between KEs. Weight of evidence for KER can be obtained by literature search (as in our case), targeted experiments, data mining, or modelling approaches. Finally, the utilization of a particular AOP within AOP-KB as the basis for an AOP network relies on a simple AOP description. It is important to mention here that our main objective was to extract and integrate relevant information from the literature/database to build a strategy to test the poten-   oli. The NM interaction with epithelial cells (bronchial or alveolar) could lead to NLRP3 (NOD-like receptor family, LRR-and pyrin domain containing 3) inflammasome activation either by lysosomal injury (can be measured by assays such as acridine orange or neutral red uptake) or membrane perturbation (can be assessed by measuring lactate dehydrogenase (LDH) release) and promote pro-inflammatory and pro-fibrotic mediator release such as IL-1β and IL-18. Caspase-1 activation is an essential component of inflammasome activation and processing of IL-1β and IL-18. Therefore, caspase-1 activity can be measured as an indicator of inflammasome activation (KE1). Subsequent IL-1β and IL-18 release can be measured by ELISA in the medium of the cell cultures to quantify pro-inflammatory and pro-fibrotic mediator release (KE2). As secreted cytokines may act along different pathways (see Fig. 4), we propose three test strategies (TS): TS1: IL-1β promotes the secretion of TGF-β1, which plays a key role in the epithelial-mesenchymal transition (EMT) (KE3). lated to lung fibrosis. In this in vitro testable AOP, we propose KE6 of AOP 173 (ECM deposition, see Fig. 3) as the AO.
Frustrated phagocytosis and lysosomal injury of MWCNTs are the key determinants of lung fibrosis initiating events (Vietti et al., 2016). When mining the literature, we have also found that high aspect ratio nanomaterials such as nanowires, nanorods and other NMs such as fumed silica and cerium oxide also induce inflammasome activation via lysosomal injury, membrane perturbation and/or frustrated phagocytosis (Wang et al., 2017). Despite the lack of information that inflammasome activation induced by these NMs potentially can lead to lung fibrosis, the downstream biological processes of the inflammasome activation induced by different NMs could be similar to that induced by MWCNTs. Therefore, we use the existing information specific for MWCNTs and propose the following strategy to test the potential of NMs to induce an AO related to lung fibrosis.
Epithelial cells are the first cells to encounter NMs once they reach the deeper parts of the lung such as bronchioles and alve-

Case study 2: Liver fibrosis
The liver is known to be one of the main target organs for ingested NMs. Therefore, we explored the AOPs for liver fibrosis presented in AOP-Wiki to generate in vitro test strategies for NMs. The schemes shown in Figure 5 both lead to liver fibrosis as follows: -AOP 144 with MIE endocytic lysosomal uptake: Endocytic lysosomal uptake (MIE) of the stressor leads to lysosomal disruption (KE1), which induces subsequent KEs at the cellular level such as mitochondrial dysfunction (KE2), cell injury and apoptosis/necrosis (KE3). Cell death leads to increased production of pro-inflammatory mediators (KE4), which attract and activate leukocytes (KE5). Activated leukocytes through molecular mediators activate hepatic stellate cells (HSCs) (KE6), which increase the accumulation of collagen (KE7) leading to extracellular matrix (ECM) alteration and AO -liver fibrosis. -AOP 38 with protein alkylation: presented with similar downstream KEs as in AOP 144 (KE3, KE6 and KE7) except that liver tissue resident macrophages release media-EMT transition can be measured by quantifying mesenchymal cell markers such as vimentin and β-catenin. These polarized epithelial cells are involved in the production of collagen (measured by collagen assay). TS2: IL-1β promotes the secretion of TGF-β1, which plays a key role in fibroblast activation and proliferation. Activated fibroblasts (KE3) are involved in the production of collagen. IL-18 is also involved in the direct activation of fibroblasts. Collagen production in exposed epithelial cells (TS1) and lung fibroblasts (TS2) can be measured as a representative in vitro AO to predict lung fibrosis in vivo.
TS3: Macrophages, the first line of defense that engulfs NMs by phagocytosis, also play a key role in the development of lung fibrosis. Upon inflammasome activation due to frustrated phagocytosis (characterized by TEM), macrophages secrete IL-1β and TNF-α, which are involved in promoting TGF-β1, which in turn activates fibroblasts (KE3) and promotes collagen production.
From the obtained results, the potency of a NM to trigger a MIE leading to AO (collagen production) via any of these pathways (TS1, 2 and 3) can be determined. sured by assays such as neutral red/acridine orange) and KE1 mitochondrial dysfunction (mitochondrial membrane potential, MMP) of hepatocytes lead to KE2 cell injury (cytotoxicity assays such as WST-1, LDH). Release of damage-associated molecular pattern (DAMP) mediators and cytokines such as TNF-α and TGF-β are involved in the direct activation of HSCs (KE3), a major source of collagen producing cells (AO) in the liver (Liu et al., 2012;Li et al., 2008). HSCs also become activated upon engulfment of DNA fragments from apoptotic hepatocytes (Li et al., 2008). To verify this, HSCs can be incubated with cell culture tors, which activate HSC (instead of mediators released by leukocytes) and with protein alkylation as MIE (Fig. 5). In the literature, it has been reported that several NMs induce lysosomal disruption and apoptosis/necrosis via lysosomal membrane permeabilization (LMP) (Stern et al., 2012). Therefore, we propose three test strategies using simple in vitro experiments with LMP as an initial KE (Fig. 6).
TS1: Hepatocytes (epithelial cells) can be used as a cell model as ingested NMs, once entering the liver via the portal vein, encounter the epithelial layer of the liver. NM-induced LMP (mea-

Fig. 6: Proposed in vitro strategy to test the potential of a NM to induce liver fibrosis
reported in nanotoxicological studies, their relevance in terms of predicting the AO is unknown. AOPs are relevant in the context of organizing the existing information to establish relationships between key biological endpoints for AO prediction. Although more than 200 AOPs are currently available in the AOP-Wiki, it is important to note that only a handful of AOPs have been formally validated or endorsed by the OECD. Furthermore, in vitro exposure models and assays that are currently being used to measure KEs need to be validated for testing of NMs. Until reaching the stage of availability of validated in vitro assays and exposure models as well as AOPs for NMs, our strategy based on existing AOPs is useful in the following contexts: (i) addressing the knowledge gaps in AOPs, (ii) early screening of NM safety assessment to prioritize animal research, and (iii) determining the influence of NM property, NM concentration, and duration of exposure in observing AO. In order to test the strategies proposed for lung and liver fibrosis, it could be worth focusing on MWCNTs because the countless number of variants differ in length, diameter, rigidness, functional groups and impurities, and because the data already points to effects on lung fibrosis in vivo for some of them (Porter et al., 2010;Mercer et al., 2011;Duke and Bonner, 2018).
Very recently, based on in vitro and in vivo studies, AOPs related to the carcinogenicity of TiO 2 NMs (Braakhuis et al., 2021) and AOPs related to steatosis, edema and fibrosis in the liver induced upon TiO 2 exposure (Brand et al., 2020;Gerloff et al., 2017) have been postulated. When analyzing these AOPs, we found that lysosomal injury or lysosomal membrane permeabilization, ROS formation, and DNA damage have been described as early KEs. These initiating KEs have also been identified in our study, indicating their potential usefulness in the characterization of NM-related MIEs. Particularly with the application of high-throughput screening and high-content assays, a large amount of NMs can be screened for their potentially hazardous nature. Application of nano-specific AOPs for human health risk assessment cannot be efficient without understanding the in vitro-in vivo correlation. Predictability of in vitro methods for in vivo AOs remains a critical issue and mainly concerns: a) NM dose-selection and dose-metrics, b) in vitro assays including cell type and assay conditions, and c) the nano-relevant reference materials including both negative and positive controls (Dobrovolskaia and McNeil, 2013). In the case of conventional chemicals, the OECD adopted and validated a certain number of in vitro assays against an in vivo response. The largest knowledge gap for NMs is related to existing in vivo data that would provide validity of an in vitro assay to in vivo AOs. The severity of the endpoint represents an important factor in determining the extent of validation that would be required. Thus, a selection and prioritization strategy would involve targeting those KEs that are shared by a number of AOPs with the most severe health outcomes and for which established in vitro assays are available. For example, both case studies presented here share a common immunotoxic response medium collected from NM-exposed hepatocyte cultures, and collagen production in exposed HSCs can be measured as a representative in vitro AO to predict liver fibrosis in vivo.
TS2: Leukocytes, upon activation, secrete TGF-β1 and TNF-α (KE3), which in turn activate HSCs (KE4) that produce collagen 2 . Therefore, leukocytes such as monocytes can be incubated with NMs, and subsequently HSCs can be incubated with the medium collected from monocytes, and collagen production in exposed HSCs can be measured as a representative in vitro AO to predict liver fibrosis in vivo.
TS3: Kupffer cells, which are the liver-resident macrophages, also play a key role in the development of liver fibrosis. Activated Kupffer cells can secrete pro-inflammatory mediators such as TNF-α and IL-6 and pro-fibrotic mediator TGF-β, which leads to HSC activation 3 (Liu et al., 2012). To verify this, Kupffer cells can be incubated with NMs and, subsequently, HSCs can be incubated with the medium collected from Kupffer cells, and collagen production in exposed HSC cells can be measured.
From the obtained results, the potency of a NM to trigger an initial KE leading to the AO (collagen production) via any of these pathways (TS1, 2 and 3) can be determined.

Discussion
The increasing number of applications of nanotechnology and the fast-growing market of nanoproducts creates an urgent need for the development of strategies to perform a fast and reliable safety testing and hazard assessment of NMs. The AOP framework represents an important regulatory relevant aid in predicting the adverse effects of NMs that may foster reduction/elimination of animal testing. In this paper, we describe a simple strategy for AOP implementation in nanotoxicology to facilitate the fast screening of NM safety and to provide an efficient aid to regulatory decision-making and the safe(r)-by-design approach to the development and use of NMs.
Chemical AOPs are not stressor-specific, and we assumed that they could be used to postulate the downstream effects of NMs if proper MIEs relevant for NMs were identified. Therefore, we systematically explored existing AOPs in AOP-Wiki using NM-relevant initiating KEs identified in the literature and proposed in vitro strategies to test the potential of a NM to induce lung and liver fibrosis as two AO case studies. In addition to these AOPs, we realized that several AOPs identified in the AOP-Wiki related to liver injury (AOP 273), kidney toxicity (AOP 257), neurotoxicity (AOP 3), and breast cancer (AOP 293) can also be potentially explored for NMs. However, more information on NM biodistribution in organs such as brain and breast are required to justify the use of these AOPs, which were proposed for conventional chemicals.
A recent review by Halappanavar et al. (2021) suggested that although cell death, membrane integrity, ROS/RNS and cytokines are among the in vitro biological endpoints that are widely assessment of KEs. Moreover, it has been well documented that many NM types, especially metal-based, interact with biological structures, affecting their fate (toxicokinetics and toxicodynamics) in biological systems (Feliu et al., 2016).
In biological media, aggregation, agglomeration, dissolution and degradation of NMs may occur, leading to the generation and co-existence of different sizes and forms (NMs, ions and complex salt forms), illustrating the need for detailed analytical tools able to pick up this mixture of chemical forms and particle types. Some of these new species may trigger MIEs and KEs that are not nano-specific. Nano-bio interactions should be assessed and used to govern selection of in vitro tests by considering specific endpoints that derive from nano-related properties and transformations driven by the biological environment present around the material. For example, lysosomal enzyme release and quantification of metal can be used to assess lysosomal dysfunction as a KE resulting from the active cellular uptake of metallic NMs and their transformation in endosomes.
In any case, no single in vitro model is sufficient to provide a comprehensive answer about safety or hazards of NMs. While validation of assays with regard to their relevance, reliability, and specificity represents the enduring need for risk governance of nanotechnology, we are still faced with the huge lack of human exposure and effects data for NMs that would foster adaptation and development of such methods.
The choice of exposure durations (short-term or long-term) and exposure concentrations is critical for the safety assessment of NMs as most of them exhibit very low acute toxicity (Annangi et al., 2016;Xi et al., 2019;Chen et al., 2016). Indeed, most in vitro studies evaluated the toxic potential of NMs after short-term exposure (24 to 72 h), while long-term and repeated low-concentration exposure studies are scarce but are extremely important as they better mimic real-life exposure (e.g., workers in production and consumers through food). Since NM safety assessment by AOP testing is in the early stages of development, use of exposure conditions and relevant in vitro models that mimic more closely the realistic exposure situation should be encouraged.
Considering that the regulatory relevant AOP networks extend and enhance the toxicity testing strategy via providing insight into the mechanism of action, they also support the development of relevant approaches for toxicity prediction including computationally-based predictive models. Hence, linking the AOP framework to in silico methods may facilitate the safety prediction of NMs. In silico models may be used to predict biological responses of potential concern for the occurrence of the AO instead of predicting the apical changes measured at the phenotypic level (Jagiello et al., 2021). However, the relevance of the response used for modelling of the eventual AO first needs to be justified. In effect, AOP-anchored predictive models (including QSAR) would be delivered. The development of mathematical models (including QSARs) as predictive tools for early KEs is now one of the long-term actions according to the OECD report (OECD, 2015). The QSAR models have been widely used for predicting for which the Nanotechnology Characterization Lab (NCL) has recommended in vitro assays with high potential for in vivo predictability (Dobrovolskaia and McNeil, 2013).
The choice of the in vitro model is the next crucial step for the successful application of in vitro testing strategies. The model must have the ability to exhibit crucial KEs upon exposure to NMs. For instance, in testing the lung fibrosis AOP, selected epithelial cells must have the ability to undergo EMT transition, whereas epithelial cells and macrophages must have the ability to release sufficient levels of pro-inflammatory and pro-fibrotic mediators. Agents that can induce these effects (positive agents) can be used to characterize the abilities of cell types to exhibit these KEs. As an example, TGF-β1, a strong promoter of EMT transition, can be used to check the ability of the selected cell type to undergo EMT transition.
However, most validated in vitro tests are based on 2D monocultures that do not reliably represent the architecture and physiology of an organ and the interactions within an organism. It has been recommended to develop and validate test systems based on a combination of cell cultures, co-cultures, tissue and tissue culture models . The AOP development for nanosafety assessment should benefit from advanced biological models such as reconstructed epithelia, 3D cultures and microfluidic-based platforms that are continuously developed in the 3Rs spirit, particularly those that allow long-term and/or lowdose exposure to better predict chronic effects (Drasler et al., 2017;Ruzycka et al., 2019;Barosova et al., 2020). Such models, developed to mimic human physiology and metabolism, hold great promise for RA of engineered NMs (Burden et al., 2021). As an example, advanced in vitro models of the human lung and liver have been used in the EU H2020 projects "Physiologically Anchored Tools for Realistic nanOmateriaL hazard aSsessment" (PATROLS 4 ) and "Smart Tools for Gauging Nano Hazards" (SmartNanoTox 5 ).
Further, one should consider the reliability, reproducibility and accessibility of the in vitro testing approach for RA of NMs with respect to nano-specific challenges, e.g., the heterogeneity of NMs, the interference of NMs with assays, and the lack of standardized protocols (Savolainen et al., 2010;Shah et al., 2014). An in vitro toolbox should include selected critical checkpoints to avoid any undesirable interactions of NMs with assay components and/or detection systems caused by NM properties or high exposure concentration, which may lead to erroneous results (Vinković Vrček et al., 2015;Hoet et al., 2013;Guadagnini et al., 2015;Ostermann et al., 2020;Kroll et al., 2011Kroll et al., , 2012Seiffert et al., 2012). Despite many studies evidencing that nano-specific properties, such as high adsorption capacity, hydrophobicity, surface charge, optical and magnetic properties, or catalytic activity may induce interferences with in vitro methods, this issue has still not been adequately considered in nanotoxicology and nanomedicine. Any false positive and false negative results caused by NM-induced interferences would undoubtedly create errors in the interpretation of in vitro velopment of computationally assisted AOP development for NMs therefore should be based on the provision of high-quality data that support establishing full AOPs and creating predictive models.
One aspect that is generally missing in biological testing is the evaluation of the uncertainty in measurements (ISO, 2008), which would facilitate the assessment of the reliability of measurements and comparison of the results obtained in different laboratories. We would therefore recommend the inclusion of measurement uncertainty as one of the important parameters for the overall data analysis.
Combined in vitro/in silico approaches may significantly facilitate AOP acceptance by all relevant stakeholder groups coming into contact with NMs with regards to the evaluation of existing information, identification of data gaps, generation of new knowledge, and iterative decision making (Ede et al., 2020). Finally, the AOP framework also finds its utility in the nano-product design and development as mechanistic aid, thereby having relevance beyond regulatory applications.

References
Allen, T. E. H., Goodman, J. M., Gutsell, S. et al. (2014 Gosens, I., Heringa, M. B. et al. (2021). Mech-the toxicity of chemicals. These models are based on the assumption that substances with similar chemical structures will have similar toxicological effects/mechanisms of action, i.e., will follow similar patterns of interaction with possible molecular targets (e.g., protein, receptor, membrane) and/or induce similar changes at the cellular level (Kubinyi, 1997).
The first reports on the use of QSAR methodology for NMs, socalled NanoQSAR, appeared approximately a decade ago (Puzyn et al., 2011). The existing NanoQSAR methods still need adjustment due to the fact that most of the models developed so far refer to the phenotypic response, mainly cytotoxicity (Puzyn et al., 2011;Pan et al., 2016;Gajewicz et al., 2015). In fact, the relevance of modelled activity for the eventual nano-related AOs is not clear. This situation reveals that the challenge for future directions is to construct AOP-informed NanoQSAR models closely connected with AOP models. Such models will link physico-chemical properties of NMs with the AOP-relevant responses. Another computational strategy being developed to foster the validation of in vitro methods for in vivo AOs is the in vitro-in vivo extrapolation (IVIVE) approach employing historical in vivo data sets as used by PATROLS and SmartNanoTox projects. As suggested (Halappanavar et al., 2020), analysis of large, heterogeneous data sets on NM toxicity by sophisticated computational models and machine learning will enable integration and description of biologically relevant information to be used for determination of causative links between KEs and KERs within AOP.
The simplest solution would be to link NM properties to the MIEs they may induce. However, this is currently not possible due to the fact that AOP development has not been focused on toxicological mechanisms relevant for NMs so far (Halappanavar et al., 2019). There is no practical knowledge on whether events and outcomes of AOPs integrated in the AOP-Wiki to date may be related to nano-related structural features (e.g., shape, size, surface coating). Secondly, there is no structured data that clearly and comprehensively describes NM-related MIEs (Halappanavar et al., 2020). Thus, additional efforts should be invested to critically review existing scientific literature and databases towards identification of all possible NM-related MIEs. Finally, lessons on biological effects of NMs learned so far suggest that biological actions of NMs are often non-specific (Halappanavar et al., 2020). More extensively defined in relation to NM-relevant AOPs are early KEs, which are the first identified biological responses essential for the occurrence of the AO to be manifested at the organism level. Responses such as inflammation, oxidative stress, or cytotoxicity are repeatedly listed as KEs in the nanotoxicology literature. All these toxicological effects may lead to the AOs that have been confirmed for NMs, e.g., lung fibrosis, lung emphysema and lung cancer, that are already described in the AOP-Wiki as AOP 173, AOP 1.25 and AOP 303, respectively (Halappanavar et al., 2020).
Recently, a novel strategy for modelling NM toxicity was proposed (Jagiello et al., 2021), which employed a nano-QSAR model able to predict transcriptomic response related to the early KEs regarding AOP 173. Such an approach could be used to show the relationship between the structural features of NMs and the AO considered by the AOP (Jagiello et al., 2021). Further de-