The role of trust in the use of artificial intelligence for chemical risk assessment.

Risk assessment of chemicals is a time-consuming process and needs to be optimized to ensure all chemicals are timely evaluated and regulated. This transition could be stimulated by valuable applications of in silico Artificial Intelligence (AI)/Machine Learning (ML) models. However, implementation of AI/ML models in risk assessment is lagging behind. Most AI/ML models are considered 'black boxes' that lack mechanistical explainability, causing risk assessors to have insufficient trust in their predictions. Here, we explore 'trust' as an essential factor towards regulatory acceptance of AI/ML models. We provide an overview of the elements of trust, including technical and beyond-technical aspects, and highlight elements that are considered most important to build trust by risk assessors. The results provide recommendations for risk assessors and computational modelers for future development of AI/ML models, including: 1) Keep models simple and interpretable; 2) Offer transparency in the data and data curation; 3) Clearly define and communicate the scope/intended purpose; 4) Define adoption criteria; 5) Make models accessible and user-friendly; 6) Demonstrate the added value in practical settings; and 7) Engage in interdisciplinary settings. These recommendations should ideally be acknowledged in future developments to stimulate trust and acceptance of AI/ML models for regulatory purposes.

Approaches to implement safe by design in early product design through combining risk assessment and Life Cycle Assessment.

The Safe by Design (SbD) concept aims to ensure the production, use and disposal of materials and products safely. While there is a growing interest in the potential of SbD to support policy commitments, such as the EU Green Deal and the Circular Economy Action Plan in Europe, methodological approaches and practical guidelines on SbD are, however, largely missing. The combined use of Life Cycle Assessment (LCA) and Risk Assessment (RA) is considered suitable to operationalize SbD over the whole life-cycle of a product. Here, we explore the potential of the combined use of LCA and RA at Technological Readiness Level (TRL) 1-6. We perform a review of the literature presenting and/or developing approaches that combine LCA and RA at early stages of product design. We identify that basic early-on-evaluations of safety (e.g., apply lifecycle thinking to assess risk hotspots, avoid use of hazardous chemicals, minimize other environmental impacts from chemicals) are more common, while more complex assessments (e.g., ex-ante LCA, control banding, predictive (eco)toxicology) require specialized expertise. The application of these simplified approaches and guidelines aims to avoid some obvious sources of risks and impacts at early stages. Critical gaps need to be addressed for wider application of SbD, including more studies in the product design context, developing tools and databases containing collated information on risk, greater collaboration between RA/LCA researchers and companies, and policy discussion on the expansion from SbD to Safe and Sustainable by Design (SSbD).

Similarity assessment of metallic nanoparticles within a risk assessment framework: A case study on metallic nanoparticles and lettuce.

Similarity assessment is one of the means of optimally using scarcely available experimental data on the fate and hazards of nanoforms (NFs) for regulatory purposes. For a set of NFs that are shown to be similar it is allowed in a regulatory context to apply the information available on any of the NFs within the group to the whole set of NFs. Obviously, a proper justification for such a similarity assessment is to be provided. Within the context of exemplifying such a justification, a case study was performed aimed at assessing the similarity of a set of spherical metallic NFs that different with regard to chemical composition (three metals) and particle size (three different sizes). The endpoints of assessment were root elongation and biomass increase of lettuce (Lactuca sativa L.) seedlings and exposure assessment was performed in order to express the actual exposure concentration in terms of time-weighted average particle concentrations. The results of the study show that for the specific endpoints assessed, chemical composition is driving NF toxicity and this is mostly due to impacts on the fate of the NFs. On the other hand, particle size of Cu NFs had a negligible impact on the dose-response relationships for the specific endpoints assessed. It is thus concluded that hazard data available on spherical Cu NF tested in our case can be used to inform on the hazards of any spherical Cu NF within the size range of 25-100 nm, but only applies for the certain endpoints. Also, toxicity data for the Cu2+-ion are suited for such a similarity assessment.

How can we justify grouping of nanoforms for hazard assessment? Concepts and tools to quantify similarity.

The risk of each nanoform (NF) of the same substance cannot be assumed to be the same, as they may vary in their physicochemical characteristics, exposure and hazard. However, neither can we justify a need for more animal testing and resources to test every NF individually. To reduce the need to test all NFs, (regulatory) information requirements may be fulfilled by grouping approaches. For such grouping to be acceptable, it is important to demonstrate similarities in physicochemical properties, toxicokinetic behaviour, and (eco)toxicological behaviour. The GRACIOUS Framework supports the grouping of NFs, by identifying suitable grouping hypotheses that describe the key similarities between different NFs. The Framework then supports the user to gather the evidence required to test these hypotheses and to subsequently assess the similarity of the NFs within the proposed group. The evidence needed to support a hypothesis is gathered by an Integrated Approach to Testing and Assessment (IATA), designed as decision trees constructed of decision nodes. Each decision node asks the questions and provides the methods needed to obtain the most relevant information. This White paper outlines existing and novel methods to assess similarity of the data generated for each decision node, either via a pairwise analysis conducted property-by-property, or by assessing multiple decision nodes simultaneously via a multidimensional analysis. For the pairwise comparison conducted property-by-property we included in this White paper: The x-fold, Bayesian and Arsinh-OWA distance algorithms performed comparably in the scoring of similarity between NF pairs. The Euclidean distance was also useful, but only with proper data transformation. The x-fold method does not standardize data, and thus produces skewed histograms, but has the advantage that it can be implemented without programming knowhow. A range of multidimensional evaluations, using for example dendrogram clustering approaches, were also investigated. Multidimensional distance metrics were demonstrated to be difficult to use in a regulatory context, but from a scientific perspective were found to offer unexpected insights into the overall similarity of very different materials. In conclusion, for regulatory purposes, a property-by-property evaluation of the data matrix is recommended to substantiate grouping, while the multidimensional approaches are considered to be tools of discovery rather than regulatory methods.

Bayesian based similarity assessment of nanomaterials to inform grouping.

Nanoforms can be manufactured in plenty of variants by differing their physicochemical properties and toxicokinetic behaviour which can affect their hazard potential. To avoid testing of each single nanomaterial and nanoform variation and subsequently save resources, grouping and read-across strategies are used to estimate groups of substances, based on carefully selected evidence, that could potentially have similar human health and environmental hazard impact. A novel computational similarity method is presented aiming to compare dose-response curves and identify sets of similar nanoforms. The suggested method estimates the statistical model that best fits the data by leveraging pairwise Bayes Factor analysis to compare pairs of curves and evaluate whether each of the nanoforms is sufficiently similar to all other nanoforms. Pairwise comparisons to benchmark materials are used to define threshold similarity values and set the criteria for identifying groups of nanoforms with comparatively similar toxicity. Applications to use case data are shown to demonstrate that the method can support grouping hypotheses linked to a certain hazard endpoint and route of exposure.

Thermochemical unification of molecular descriptors to predict radical hydrogen abstraction with low computational cost.

Chemistry describes transformation of matter with reaction equations and corresponding rate constants. However, accurate rate constants are not always easy to get. Here we focus on radical oxidation reactions. Analysis of over 500 published rate constants of hydroxyl radicals led us to hypothesize that a modified linear free-energy relationship (LFER) could be used to predict rate constants speedily, reliably and accurately. LFERs correlate the Gibbs activation-energy with the Gibbs energy of reaction. We calculated the latter as the sum of one-electron transfer and, if appropriate, proton transfer. We parametrized specific transition state effects to orbital delocalizability and the polarity of the reactant. The calculation time for 500 reactions is less than 8 hours on a standard desktop-PC. Rate constants were also calculated for hydrogen and methyl radicals; these controls show that the predictions are applicable to a broader set of oxidizing radicals. An accuracy of 30-40% (standard deviation) with reference to reported experimental values was found suitable for the screening of complex chemical systems for possibly relevant reactions. In particular, potentially relevant reactions can be singled out and scrutinized in detail when prioritizing chemicals for environmental risk assessment.

Toward harmonizing ecotoxicity characterization in life cycle impact assessment.

Ecosystem quality is an important area of protection in life cycle impact assessment (LCIA). Chemical pollution has adverse impacts on ecosystems on a global scale. To improve methods for assessing ecosystem impacts, the Life Cycle Initiative hosted by the United Nations Environment Programme established a task force to evaluate the state-of-the-science in modeling chemical exposure of organisms and the resulting ecotoxicological effects for use in LCIA. The outcome of the task force work will be global guidance and harmonization by recommending changes to the existing practice of exposure and effect modeling in ecotoxicity characterization. These changes will reflect the current science and ensure the stability of recommended practice. Recommendations must work within the needs of LCIA in terms of 1) operating on information from any inventory reporting chemical emissions with limited spatiotemporal information, 2) applying best estimates rather than conservative assumptions to ensure unbiased comparison with results for other impact categories, and 3) yielding results that are additive across substances and life cycle stages and that will allow a quantitative expression of damage to the exposed ecosystem. We describe the current framework and discuss research questions identified in a roadmap. Primary research questions relate to the approach toward ecotoxicological effect assessment, the need to clarify the method's scope and interpretation of its results, the need to consider additional environmental compartments and impact pathways, and the relevance of effect metrics other than the currently applied geometric mean of toxicity effect data across species. Because they often dominate ecotoxicity results in LCIA, we give metals a special focus, including consideration of their possible essentiality and changes in environmental bioavailability. We conclude with a summary of key questions along with preliminary recommendations to address them as well as open questions that require additional research efforts. Environ Toxicol Chem 2018;37:2955-2971. © 2018 SETAC.

Dissipative particle dynamic simulation and experimental assessment of the impacts of humic substances on aqueous aggregation and dispersion of engineered nanoparticles.

Comprehensive experimental quantification and mapping of the aggregation and dispersion state of engineered nanoparticles (NPs) in the presence of humic substances is a great challenge. Dissipative particle dynamic (DPD) simulation was adopted to investigate the aggregation and dispersion mechanisms of NPs in the presence of a humic substance analog. Twelve different types of NPs including 2 metal-based NPs, 7 metal oxide-based NPs, and 3 carbon-based NPs in pure water (pH 3.0) and algae medium (pH 8.0) in the presence of a humic substance analogy were selected for experimental verification of the DPD simulation results. In agreement with results obtained with dynamic light scattering and phase analysis light scattering techniques, the simulations demonstrated that the presence of humic substances reduced the aggregation extent of the NPs. The DPD simulations showed that the stability and dispersity of the NPs increased first, and then decreased with increasing concentrations of humic substances. Moreover, there existed a concentration of humic substances where the NPs became more stable and more dispersed, which was experimentally verified in the case of all the NPs in the pure water and in the algae medium. Furthermore, theory and simulation indicate that both hydrophobic and hydrogen interaction play an important role in controlling the formation of NP aggregates in the presence of humic substances. Electrostatic interaction and steric repulsion are the main mechanisms underlying the effects of humic substances on the aqueous dispersion stability of NPs. Environ Toxicol Chem 2018;37:1024-1031. © 2017 SETAC.

Environmental Risk Assessment Strategy for Nanomaterials.

An Environmental Risk Assessment (ERA) for nanomaterials (NMs) is outlined in this paper. Contrary to other recent papers on the subject, the main data requirements, models and advancement within each of the four risk assessment domains are described, i.e., in the: (i) materials, (ii) release, fate and exposure, (iii) hazard and (iv) risk characterisation domains. The material, which is obviously the foundation for any risk assessment, should be described according to the legislatively required characterisation data. Characterisation data will also be used at various levels within the ERA, e.g., exposure modelling. The release, fate and exposure data and models cover the input for environmental distribution models in order to identify the potential (PES) and relevant exposure scenarios (RES) and, subsequently, the possible release routes, both with regard to which compartment(s) NMs are distributed in line with the factors determining the fate within environmental compartment. The initial outcome in the risk characterisation will be a generic Predicted Environmental Concentration (PEC), but a refined PEC can be obtained by applying specific exposure models for relevant media. The hazard information covers a variety of representative, relevant and reliable organisms and/or functions, relevant for the RES and enabling a hazard characterisation. The initial outcome will be hazard characterisation in test systems allowing estimating a Predicted No-Effect concentration (PNEC), either based on uncertainty factors or on a NM adapted version of the Species Sensitivity Distributions approach. The risk characterisation will either be based on a deterministic risk ratio approach (i.e., PEC/PNEC) or an overlay of probability distributions, i.e., exposure and hazard distributions, using the nano relevant models.

Setting the stage for debating the roles of risk assessment and life-cycle assessment of engineered nanomaterials.

Although technological and environmental benefits are important stimuli for nanotechnology development, these technologies have been contested from an environmental point of view. The steady growth of applications of engineered nanomaterials has heated up the debate on quantifying the environmental repercussions. The two main scientific methods to address these environmental repercussions are risk assessment and life-cycle assessment. The strengths and weaknesses of each of these methods, and the relation between them, have been a topic of debate in the world of traditional chemistry for over two decades. Here we review recent developments in this debate in general and for the emerging field of nanomaterials specifically. We discuss the pros and cons of four schools of thought for combining and integrating risk assessment and life-cycle assessment and conclude with a plea for action.

Current Knowledge on the Use of Computational Toxicology in Hazard Assessment of Metallic Engineered Nanomaterials.

As listed by the European Chemicals Agency, the three elements in evaluating the hazards of engineered nanomaterials (ENMs) include the integration and evaluation of toxicity data, categorization and labeling of ENMs, and derivation of hazard threshold levels for human health and the environment. Assessing the hazards of ENMs solely based on laboratory tests is time-consuming, resource intensive, and constrained by ethical considerations. The adoption of computational toxicology into this task has recently become a priority. Alternative approaches such as (quantitative) structure-activity relationships ((Q)SAR) and read-across are of significant help in predicting nanotoxicity and filling data gaps, and in classifying the hazards of ENMs to individual species. Thereupon, the species sensitivity distribution (SSD) approach is able to serve the establishment of ENM hazard thresholds sufficiently protecting the ecosystem. This article critically reviews the current knowledge on the development of in silico models in predicting and classifying the hazard of metallic ENMs, and the development of SSDs for metallic ENMs. Further discussion includes the significance of well-curated experimental datasets and the interpretation of toxicity mechanisms of metallic ENMs based on reported models. An outlook is also given on future directions of research in this frontier.

Assessment and prediction of joint algal toxicity of binary mixtures of graphene and ionic liquids.

Graphene and ionic liquids (ILs) released into the environment will interact with each other. So far however, the risks associated with the concurrent exposure of biota to graphene and ILs in the environment have received little attention. The research reported here focused on observing and predicting the joint toxicity effects in the green alga Scenedesmus obliquus exposed to binary mixtures of intrinsic graphene (iG)/graphene oxide (GO) and five ILs of varying anionic and cationic types. The isolated ILs in the binary mixtures were the main contributors to toxicity. The binary GO-IL mixtures resulted in more severe joint toxicity than the binary iG-IL mixtures, irrespective of mixture ratios. The mechanism of the joint toxicity may be associated with the adsorption capability of the graphenes for the ILs, the dispersion stability of the graphenes in aquatic media, and modulation of the binary mixtures-induced oxidative stress. A toxic unit assessment showed that the graphene and IL toxicities were additive at low concentration of the mixtures but antagonistic at high concentration of the mixtures. Predictions made using the concentration addition and independent action models were close to the observed joint toxicities regardless of mixture types and mixture ratios. These findings provide new insights that are of use in the risk assessment of mixtures of engineered nanoparticles and other environmentally relevant contaminants.

Regulatory ecotoxicity testing of nanomaterials - proposed modifications of OECD test guidelines based on laboratory experience with silver and titanium dioxide nanoparticles.

Regulatory ecotoxicity testing of chemicals is of societal importance and a large effort is undertaken at the OECD to ensure that OECD test guidelines (TGs) for nanomaterials (NMs) are available. Significant progress to support the adaptation of selected TGs to NMs was achieved in the context of the project MARINA ( http://www.marina-fp7.eu/ ) funded within the 7th European Framework Program. Eight OECD TGs were adapted based on the testing of at least one ion-releasing NM (Ag) and two inert NMs (TiO2). With the materials applied, two main variants of NMs (ion releasing vs. inert NMs) were addressed. As the modifications of the test guidelines refer to general test topics (e.g. test duration or measuring principle), we assume that the described approaches and modifications will be suitable for the testing of further NMs with other chemical compositions. Firm proposals for modification of protocols with scientific justification(s) are presented for the following tests: growth inhibition using the green algae Raphidocelis subcapitata (formerly: Pseudokirchneriella subcapitata; TG 201), acute toxicity with the crustacean Daphnia magna (TG 202), development toxicity with the fish Danio rerio (TG 210), reproduction of the sediment-living worm Lumbriculus variegatus (TG 225), activity of soil microflora (TGs 216, 217), and reproduction of the invertebrates (Enchytraeus crypticus, Eisenia fetida, TGs 220, 222). Additionally, test descriptions for two further test systems (root elongation of plants in hydroponic culture; test on fish cells) are presented. Ecotoxicological data obtained with the modified test guidelines for TiO2 NMs and Ag NM and detailed method descriptions are available.

Frameworks and tools for risk assessment of manufactured nanomaterials.

Commercialization of nanotechnologies entails a regulatory requirement for understanding their environmental, health and safety (EHS) risks. Today we face challenges to assess these risks, which emerge from uncertainties around the interactions of manufactured nanomaterials (MNs) with humans and the environment. In order to reduce these uncertainties, it is necessary to generate sound scientific data on hazard and exposure by means of relevant frameworks and tools. The development of such approaches to facilitate the risk assessment (RA) of MNs has become a dynamic area of research. The aim of this paper was to review and critically analyse these approaches against a set of relevant criteria. The analysis concluded that none of the reviewed frameworks were able to fulfill all evaluation criteria. Many of the existing modelling tools are designed to provide screening-level assessments rather than to support regulatory RA and risk management. Nevertheless, there is a tendency towards developing more quantitative, higher-tier models, capable of incorporating uncertainty into their analyses. There is also a trend towards developing validated experimental protocols for material identification and hazard testing, reproducible across laboratories. These tools could enable a shift from a costly case-by-case RA of MNs towards a targeted, flexible and efficient process, based on grouping and read-across strategies and compliant with the 3R (Replacement, Reduction, Refinement) principles. In order to facilitate this process, it is important to transform the current efforts on developing databases and computational models into creating an integrated data and tools infrastructure to support the risk assessment and management of MNs.

Dose metrics assessment for differently shaped and sized metal-based nanoparticles.

The mean primary diameter (d) of nanoparticles (NPs) is commonly used as the best basis to assess the dose metric for expressing the toxicity of spherical NPs. However, d bears no relevance for nonspherical NPs. In the present study, the mean surface area to volume ratio, applicable to both spherical and nonspherical NPs, was used to replace d given the aim of obtaining the best dose metric (volume, surface area, or number) of differently shaped metallic NPs in vivo (9 organisms) and in vitro (6 mammalian cell lines). The slope of the curves obtained by relating the total particle number of NPs at various effect concentrations to the mean surface area to volume ratio was subsequently used to deduce the best dose metric. For the majority of the organisms studied, it was found that NP volume is the most appropriate dose metric, independent of the composition of the NPs tested. For 3 organisms exposed to Ag NPs, however, surface area was found to be the best dose metric. It is therefore concluded that the optimum dose metric depends on organism and NP properties. Environ Toxicol Chem 2016;35:2466-2473. © 2016 SETAC.

The way forward for risk assessment of nanomaterials in solid media.

There is an urgent need for sufficient knowledge to allow reliable assessment of the risks associated with engineered nanomaterials (ENPs). Significant advances in basic understanding of nano safety have been made, but there is still no clear systematic basis for risk-related research, and major uncertainties remain in the absence of uniform procedures. The following papers provide the guidance on how to proceed within the area of fate and hazard assessment, and how this links into grouping, testing and risk assessment of nanomaterials. This guidance is coupled with an industrial view on the most important research areas for nanomaterials.

The MARINA Risk Assessment Strategy: A Flexible Strategy for Efficient Information Collection and Risk Assessment of Nanomaterials.

An engineered nanomaterial (ENM) may actually consist of a population of primary particles, aggregates and agglomerates of various sizes. Furthermore, their physico-chemical characteristics may change during the various life-cycle stages. It will probably not be feasible to test all varieties of all ENMs for possible health and environmental risks. There is therefore a need to further develop the approaches for risk assessment of ENMs. Within the EU FP7 project Managing Risks of Nanoparticles (MARINA) a two-phase risk assessment strategy has been developed. In Phase 1 (Problem framing) a base set of information is considered, relevant exposure scenarios (RESs) are identified and the scope for Phase 2 (Risk assessment) is established. The relevance of an RES is indicated by information on exposure, fate/kinetics and/or hazard; these three domains are included as separate pillars that contain specific tools. Phase 2 consists of an iterative process of risk characterization, identification of data needs and integrated collection and evaluation of data on the three domains, until sufficient information is obtained to conclude on possible risks in a RES. Only data are generated that are considered to be needed for the purpose of risk assessment. A fourth pillar, risk characterization, is defined and it contains risk assessment tools. This strategy describes a flexible and efficient approach for data collection and risk assessment which is essential to ensure safety of ENMs. Further developments are needed to provide guidance and make the MARINA Risk Assessment Strategy operational. Case studies will be needed to refine the strategy.

Grouping and Read-Across Approaches for Risk Assessment of Nanomaterials.

Physicochemical properties of chemicals affect their exposure, toxicokinetics/fate and hazard, and for nanomaterials, the variation of these properties results in a wide variety of materials with potentially different risks. To limit the amount of testing for risk assessment, the information gathering process for nanomaterials needs to be efficient. At the same time, sufficient information to assess the safety of human health and the environment should be available for each nanomaterial. Grouping and read-across approaches can be utilised to meet these goals. This article presents different possible applications of grouping and read-across for nanomaterials within the broader perspective of the MARINA Risk Assessment Strategy (RAS), as developed in the EU FP7 project MARINA. Firstly, nanomaterials can be grouped based on limited variation in physicochemical properties to subsequently design an efficient testing strategy that covers the entire group. Secondly, knowledge about exposure, toxicokinetics/fate or hazard, for example via properties such as dissolution rate, aspect ratio, chemical (non-)activity, can be used to organise similar materials in generic groups to frame issues that need further attention, or potentially to read-across. Thirdly, when data related to specific endpoints is required, read-across can be considered, using data from a source material for the target nanomaterial. Read-across could be based on a scientifically sound justification that exposure, distribution to the target (fate/toxicokinetics) and hazard of the target material are similar to, or less than, the source material. These grouping and read-across approaches pave the way for better use of available information on nanomaterials and are flexible enough to allow future adaptations related to scientific developments.

Incorporating bioavailability into toxicity assessment of Cu-Ni, Cu-Cd, and Ni-Cd mixtures with the extended biotic ligand model and the WHAM-F(tox) approach.

There are only a limited number of studies that have developed appropriate models which incorporate bioavailability to estimate mixture toxicity. Here, we explored the applicability of the extended biotic ligand model (BLM) and the WHAM-F(tox) approach for predicting and interpreting mixture toxicity, with the assumption that interactions between metal ions obey the BLM theory. Seedlings of lettuce Lactuca sativa were exposed to metal mixtures (Cu-Ni, Cu-Cd, and Ni-Cd) contained in hydroponic solutions for 4 days. Inhibition to root elongation was the endpoint used to quantify the toxic response. Assuming that metal ions compete with each other for binding at a single biotic ligand, the extended BLM succeeded in predicting toxicity of three mixtures to lettuce, with more than 82% of toxicity variation explained. There were no significant differences in the values of f(mix50) (i.e., the overall amounts of metal ions bound to the biotic ligand inducing 50% effect) for the three mixture combinations, showing the possibility of extrapolating these values to other binary metal combinations. The WHAM-F(tox) approach showed a similar level of precision in estimating mixture toxicity while requiring fewer parameters than the BLM-f(mix) model. External validation of the WHAM-F(tox) approach using literature data showed its applicability for other species and other mixtures. The WHAM-F(tox) model is suitable for delineating mixture effects where the extended BLM also applies. Therefore, in case of lower data availability, we recommend the lower parameterized WHAM-F(tox) as an effective approach to incorporate bioavailability in quantifying mixture toxicity.