Application of Bayesian networks for hazard ranking of nanomaterials to support human health risk assessment

Bioinformatics and machine learning to support nanomaterial grouping

Nanomaterials (NMs) offer plenty of novel functionalities. Moreover, their physicochemical properties can be fine-tuned to meet the needs of specific applications, leading to virtually unlimited numbers of NM variants. Hence, efficient hazard and risk assessment strategies building on New Approach Methodologies (NAMs) become indispensable. Indeed, the design, the development and implementation of NAMs has been a major topic in a substantial number of research projects. One of the promising strategies that can help to deal with the high number of NMs variants is grouping and read-across. Based on demonstrated structural and physicochemical similarity, NMs can be grouped and assessed together. Within an established NM group, read-across may be performed to fill in data gaps for data-poor variants using existing data for NMs within the group. Establishing a group requires a sound justification, usually based on a grouping hypothesis that links specific physicochemical properties to well-defined hazard endpoints. However, for NMs these interrelationships are only beginning to be understood. The aim of this review is to demonstrate the power of bioinformatics with a specific focus on Machine Learning (ML) approaches to unravel the NM Modes-of-Action (MoA) and identify the properties that are relevant to specific hazards, in support of grouping strategies. This review emphasizes the following messages: 1) ML supports identification of the most relevant properties contributing to specific hazards; 2) ML supports analysis of large omics datasets and identification of MoA patterns in support of hypothesis formulation in grouping approaches; 3) omics approaches are useful for shifting away from consideration of single endpoints towards a more mechanistic understanding across multiple endpoints gained from one experiment; and 4) approaches from other fields of Artificial Intelligence (AI) like Natural Language Processing or image analysis may support automated extraction and interlinkage of information related to NM toxicity. Here, existing ML models for predicting NM toxicity and for analyzing omics data in support of NM grouping are reviewed. Various challenges related to building robust models in the field of nanotoxicology exist and are also discussed.

A data reusability assessment in the nanosafety domain based on the NSDRA framework followed by an exploratory quantitative structure activity relationships (QSAR) modeling targeting cellular viability

INTRODUCTION

The current effort towards the digital transformation across multiple scientific domains requires data that is Findable, Accessible, Interoperable and Reusable (FAIR). In addition to the FAIR data, what is required for the application of computational tools, such as Quantitative Structure Activity Relationships (QSARs), is a sufficient data volume and the ability to merge sources into homogeneous digital assets. In the nanosafety domain there is a lack of FAIR available metadata.

METHODOLOGY

To address this challenge, we utilized 34 datasets from the nanosafety domain by exploiting the NanoSafety Data Reusability Assessment (NSDRA) framework, which allowed the annotation and assessment of dataset's reusability. From the framework's application results, eight datasets targeting the same endpoint (i.e. numerical cellular viability) were selected, processed and merged to test several hypothesis including universal versus nanogroup-specific QSAR models (metal oxide and nanotubes), and regression versus classification Machine Learning (ML) algorithms.

RESULTS

Universal regression and classification QSARs reached an 0.86 R2 and 0.92 accuracy, respectively, for the test set. Nanogroup-specific regression models reached 0.88 R2 for nanotubes test set followed by metal oxide (0.78). Nanogroup-specific classification models reached 0.99 accuracy for nanotubes test set, followed by metal oxide (0.91). Feature importance revealed different patterns depending on the dataset with common influential features including core size, exposure conditions and toxicological assay. Even in the case where the available experimental knowledge was merged, the models still failed to correctly predict the outputs of an unseen dataset, revealing the cumbersome conundrum of scientific reproducibility in realistic applications of QSAR for nanosafety. To harness the full potential of computational tools and ensure their long-term applications, embracing FAIR data practices is imperative in driving the development of responsible QSAR models.

CONCLUSIONS

This study reveals that the digitalization of nanosafety knowledge in a reproducible manner has a long way towards its successful pragmatic implementation. The workflow carried out in the study shows a promising approach to increase the FAIRness across all the elements of computational studies, from dataset's annotation, selection, merging to FAIR modeling reporting. This has significant implications for future research as it provides an example of how to utilize and report different tools available in the nanosafety knowledge system, while increasing the transparency of the results. One of the main benefits of this workflow is that it promotes data sharing and reuse, which is essential for advancing scientific knowledge by making data and metadata FAIR compliant. In addition, the increased transparency and reproducibility of the results can enhance the trustworthiness of the computational findings.

Converting Nanotoxicity Data to Information Using Artificial Intelligence and Simulation

Decades of nanotoxicology research have generated extensive and diverse data sets. However, data is not equal to information. The question is how to extract critical information buried in vast data streams. Here we show that artificial intelligence (AI) and molecular simulation play key roles in transforming nanotoxicity data into critical information, i.e., constructing the quantitative nanostructure (physicochemical properties)-toxicity relationships, and elucidating the toxicity-related molecular mechanisms. For AI and molecular simulation to realize their full impacts in this mission, several obstacles must be overcome. These include the paucity of high-quality nanomaterials (NMs) and standardized nanotoxicity data, the lack of model-friendly databases, the scarcity of specific and universal nanodescriptors, and the inability to simulate NMs at realistic spatial and temporal scales. This review provides a comprehensive and representative, but not exhaustive, summary of the current capability gaps and tools required to fill these formidable gaps. Specifically, we discuss the applications of AI and molecular simulation, which can address the large-scale data challenge for nanotoxicology research. The need for model-friendly nanotoxicity databases, powerful nanodescriptors, new modeling approaches, molecular mechanism analysis, and design of the next-generation NMs are also critically discussed. Finally, we provide a perspective on future trends and challenges.

Data-Driven Quantitative Intrinsic Hazard Criteria for Nanoproduct Development in a Safe-by-Design Paradigm: A Case Study of Silver Nanoforms

The current European (EU) policies, that is, the Green Deal, envisage safe and sustainable practices for chemicals, which include nanoforms (NFs), at the earliest stages of innovation. A theoretically safe and sustainable by design (SSbD) framework has been established from EU collaborative efforts toward the definition of quantitative criteria in each SSbD dimension, namely, the human and environmental safety dimension and the environmental, social, and economic sustainability dimensions. In this study, we target the safety dimension, and we demonstrate the journey toward quantitative intrinsic hazard criteria derived from findable, accessible, interoperable, and reusable data. Data were curated and merged for the development of new approach methodologies, that is, quantitative structure-activity relationship models based on regression and classification machine learning algorithms, with the intent to predict a hazard class. The models utilize system (i.e., hydrodynamic size and polydispersity index) and non-system (i.e., elemental composition and core size)-dependent nanoscale features in combination with biological in vitro attributes and experimental conditions for various silver NFs, functional antimicrobial textiles, and cosmetics applications. In a second step, interpretable rules (criteria) followed by a certainty factor were obtained by exploiting a Bayesian network structure crafted by expert reasoning. The probabilistic model shows a predictive capability of ≈78% (average accuracy across all hazard classes). In this work, we show how we shifted from the conceptualization of the SSbD framework toward the realistic implementation with pragmatic instances. This study reveals (i) quantitative intrinsic hazard criteria to be considered in the safety aspects during synthesis stage, (ii) the challenges within, and (iii) the future directions for the generation and distillation of such criteria that can feed SSbD paradigms. Specifically, the criteria can guide material engineers to synthesize NFs that are inherently safer from alternative nanoformulations, at the earliest stages of innovation, while the models enable a fast and cost-efficient in silico toxicological screening of previously synthesized and hypothetical scenarios of yet-to-be synthesized NFs.

Reducing the question burden of patient reported outcome measures using Bayesian networks

Patient Reported Outcome Measures (PROMs) are questionnaires completed by patients about aspects of their health status. They are a vital part of learning health systems as they are the primary source of information about important outcomes that are best assessed by patients such as pain, disability, anxiety and depression. The volume of questions can easily become burdensome. Previous techniques reduced this burden by dynamically selecting questions from question item banks which are specifically built for different latent constructs being measured. These techniques analyzed the information function between each question in the item bank and the measured construct based on item response theory then used this information function to dynamically select questions by computerized adaptive testing. Here we extend those ideas by using Bayesian Networks (BNs) to enable Computerized Adaptive Testing (CAT) for efficient and accurate question selection on widely-used existing PROMs. BNs offer more comprehensive probabilistic models of the connections between different PROM questions, allowing the use of information theoretic techniques to select the most informative questions. We tested our methods using five clinical PROM datasets, demonstrating that answering a small subset of questions selected with CAT has similar predictions and error to answering all questions in the PROM BN. Our results show that answering 30% - 75% questions selected with CAT had an average area under the receiver operating characteristic curve (AUC) of 0.92 (min: 0.8 - max: 0.98) for predicting the measured constructs. BNs outperformed alternative CAT approaches with a 5% (min: 0.01% - max: 9%) average increase in the accuracy of predicting the responses to unanswered question items.

Predicting the cytotoxicity of nanomaterials through explainable, extreme gradient boosting

Nanoparticles (NPs) are a wide class of materials currently used in several industrial and biomedical applications. Due to their small size (1-100 nm), NPs can easily enter the human body, inducing tissue damage. NP toxicity depends on physical and chemical NP properties (e.g., size, charge and surface area) in ways and magnitudes that are still unknown. We assess the average as well as the individual importance of NP atomic descriptors, along with chemical properties and experimental conditions, in determining cytotoxicity endpoints for several nanomaterials. We employ a multicenter cytotoxicity nanomaterial database (12 different materials with first and second dimensions ranging between 2.70 and 81.2 nm and between 4.10 and 4048 nm, respectively). We develop a regressor model based on extreme gradient boosting with hyperparameter optimization. We employ Shapley additive explanations to obtain good cytotoxicity prediction performance. Model performances are quantified as statistically significant Spearman correlations between the true and predicted values, ranging from 0.5 to 0.7. Our results show that i) size in situ and surface areas larger than 200 nm and 50 m2/g, respectively, ii) primary particles smaller than 20 nm; iii) irregular (i.e., not spherical) shapes and iv) positive Z-potentials contribute the most to the prediction of NP cytotoxicity, especially if lactate dehydrogenase (LDH) assays are employed for short experimental times. These results were moderately stable across toxicity endpoints, although some degree of variability emerged across dose quantification methods, confirming the complexity of nano-bio interactions and the need for large, systematic experimental characterization to reach a safer-by-design approach.

Experimental and Computational Nanotoxicology-Complementary Approaches for Nanomaterial Hazard Assessment

The growing development and applications of nanomaterials lead to an increasing release of these materials in the environment. The adverse effects they may elicit on ecosystems or human health are not always fully characterized. Such potential toxicity must be carefully assessed with the underlying mechanisms elucidated. To that purpose, different approaches can be used. First, experimental toxicology consisting of conducting in vitro or in vivo experiments (including clinical studies) can be used to evaluate the nanomaterial hazard. It can rely on variable models (more or less complex), allowing the investigation of different biological endpoints. The respective advantages and limitations of in vitro and in vivo models are discussed as well as some issues associated with experimental nanotoxicology. Perspectives of future developments in the field are also proposed. Second, computational nanotoxicology, i.e., in silico approaches, can be used to predict nanomaterial toxicity. In this context, we describe the general principles, advantages, and limitations especially of quantitative structure–activity relationship (QSAR) models and grouping/read-across approaches. The aim of this review is to provide an overview of these different approaches based on examples and highlight their complementarity.

Quantifying uncertainty in dose-response screenings of nanoparticles: a Bayesian data analysis

Fitting theoretical models to experimental data for dose-response screenings of nanoparticles yields values of several hazard metrics that can support risk management. In this paper, we describe a Bayesian approach to the analysis of dose-response data for nanoparticles that takes into account multiple sources of uncertainty. Specifically, we develop a Bayesian model for the analysis of data for the cytotoxicity of ZnO nanoparticles that follow the log-logistic equation. This model reproduces the unequal variance across doses observed in the experimental data, incorporates information about the sensitivity of the cytotoxicity assay used (i.e. resazurin), and complements experimental data with historical information about the system. The model determines probability distributions for multiple values of toxicity potency (EC₅₀), and exponential decay (the slope s); these distributions provide a direct measure of uncertainty in terms of probabilistic credibility intervals. By substituting these distributions in the log-logistic equation, we determine upper and lower limits of the benchmark dose (BMD), corresponding to upper and lower limits of credibility intervals with 95% probability given the experimental data, multiple sources of uncertainty, and historical information. In view of a reduction of costs and time of dose-response screenings, we use the Bayesian model for the cytotoxicity of ZnO nanoparticles to identify the experimental design that uses the minimum number of data while reducing uncertainty in the estimation of both fitting parameters and BMD.

Bayesian Networks modeling of diarrhetic shellfish poisoning in Mytilus edulis harvested in Bantry Bay, Ireland

Diarrhetic Shellfish Poisoning (DSP) results from the human consumption of contaminated shellfish with marine biotoxins, which are produced by some species of marine dinoflagellates, mainly belonging to the genus Dinophysis. Shellfish contamination with marine biotoxins not only pose a threat to human health, but also lead to financial loss to aquaculture operations from the temporary closure of production areas when toxin concentrations exceed regulatory levels. In this study, we developed a Bayesian Network (BN) model for forecasting the short-term variations of DSP toxins in blue mussels (Mytilus edulis) from Bantry Bay, Southwest Ireland. Data inputs to a BN model from 10 production sites in Bantry Bay included plankton cell densities in sea water, DSP toxin concentration in mussels and sea surface temperature. The model was trained with data from 2014 to 2018, and validated with data of 2019. Validation consisted of predicting the DSP toxin concentration at one production site using the model parameters from the other locations as input values. Model validation showed that the prediction accuracy was higher than 86%. Sensitivity analysis indicated that in general, DSP toxin concentration was more relevant than plankton abundance. This initial work has demonstrated the usefulness of BN modeling as an approach to short term forecasting. Further work is ongoing to use the model for scenario testing and to increase the number of environmental parameters used as inputs to the model.

Health and environmental safety of nanomaterials: O Data, Where Art Thou?

Nanotechnology keeps drawing attention due to the great tunable properties of nanomaterials in comparison to their bulk conventional materials. The growth of nanotechnology in combination with the digitization era has led to an increased need of safety related data. In addition to safety, new data-driven paradigms on safe and sustainable by design materials are stressing the necessity of data even more. Data is a fundamental asset to the scientific community in studying and analysing the entire life-cycle of nanomaterials. Unfortunately, data exist in a scattered fashion, in different sources and formats. To our knowledge, there is no study focusing on aspects of actual data-structure knowledge that exists in literature and databases. The purpose of this review research is to transparently and comprehensively, display to the nanoscience community the datasets readily available for machine learning purposes making it convenient and more efficient for the next users such as modellers or data curators to retrieve information. We systematically recorded the features and descriptors available in the datasets and provide synopsised information on their ranges, forms and metrics in the supplementary material.

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.

Application of machine learning to the monitoring and prediction of food safety: A review

Machine learning (ML) has proven to be a useful technology for data analysis and modeling in a wide variety of domains, including food science and engineering. The use of ML models for the monitoring and prediction of food safety is growing in recent years. Currently, several studies have reviewed ML applications on foodborne disease and deep learning applications on food. This article presents a literature review on ML applications for monitoring and predicting food safety. The paper summarizes and categorizes ML applications in this domain, categorizes and discusses data types used for ML modeling, and provides suggestions for data sources and input variables for future ML applications. The review is based on three scientific literature databases: Scopus, CAB Abstracts, and IEEE. It includes studies that were published in English in the period from January 1, 2011 to April 1, 2021. Results show that most studies applied Bayesian networks, Neural networks, or Support vector machines. Of the various ML models reviewed, all relevant studies showed high prediction accuracy by the validation process. Based on the ML applications, this article identifies several avenues for future studies applying ML models for the monitoring and prediction of food safety, in addition to providing suggestions for data sources and input variables.

Meta-analysis of cellular toxicity for graphene via data-mining the literature and machine learning

Since graphene is currently incorporated into various consumer products and used in a variety of applications, determining the relationships between the physicochemical properties of graphene and its toxicity is critical for conducting environmental and health risk analyses. Data from the literature suggest that exposure to graphene may result in cytotoxicity. However, existing graphene toxicity data are complex and heterogeneous, making it difficult to conduct risk assessments. Here, we conducted a meta-analysis of published data on the cytotoxicity of graphene based on 792 publications, including 986 cell viability data points, 762 half maximal inhibitory concentration (IC₅₀) data points, and 100 lactate dehydrogenase (LDH) release data points. Models to predict graphene cytotoxicity were then developed based on cell viability, IC₅₀, and LDH release as toxicity endpoints using random forests learning algorithms. The most influential attributes influencing graphene cytotoxicity were revealed to be exposure dose and detection method for cell viability, diameter and surface modification for IC₅₀, and detection method and organ source for LDH release. The meta-analysis produced three sets of key attributes for the three abovementioned toxicity endpoints that can be used in future studies of graphene toxicity. The findings indicate that rigorous data mining protocols can be combined with suitable machine learning tools to develop models with good predictive power and accuracy. The results also provide guidance for the design of safe graphene materials.

A semi-quantitative risk ranking of potential human exposure to engineered nanoparticles (ENPs) in Europe

Large quantities of engineered nanoparticles (ENPs) have emerged on the European market with the rapid development of nanotechnology, however knowledge of potential health risks to humans remains in its infancy. The ENP safety issue is of pressing concern as their novel physicochemical characteristics have been illustrated compared to other bulk-form counterparts. Therefore, it is critical to carry out a comprehensive risk assessment for ENPs to guide risk management in industrial sectors. Based on current data availability, a risk ranking model is developed in accordance with the European Chemicals Agency (ECHA) advice for ENP risk assessment. In this study a Quantity, Exposure, Hazard (QEH) risk scoring model was adopted for characterizing both quantitative and qualitative data, including potential exposure pathways and hazard information. Scores were assigned to quantities of ENPs used in consumer products, intake likelihoods (oral, inhalation, and dermal intake), and hazard potential. Exposure through environmental routes and through consumer products are regarded as significant potential exposure routes. This model prioritized ENPs used in Europe according to human health risk potential. Nano-titanium dioxide (TiO₂) ranked the highest, resulting from exposure through consumer products. Silver nanoparticles (AgNP), as the second most critical ENP, is of most concern in terms of the risk from environmental sinks. Regarding the compartmentalization of total ENP risks to humans, the consumption of consumer products with nano-ingredients, especially nano-TiO₂, nano-silicon dioxide (SiO₂), and AgNP, constitutes the majority of the QEH risk index. The inadequacy of ENP risk management procedures is highlighted, not only during manufacturing, but also during nanomaterial waste disposal processes from marketplace through to the environment. Current risk assessments are based upon recent knowledge of the ENP class as novel pollutants, highlighting the need for further quantification of underlying risks as data emerges.

NanoTox: Development of a Parsimonious In Silico Model for Toxicity Assessment of Metal-Oxide Nanoparticles Using Physicochemical Features

Metal-oxide nanoparticles find widespread applications in mundane life today, and cost-effective evaluation of their cytotoxicity and ecotoxicity is essential for sustainable progress. Machine learning models use existing experimental data and learn quantitative feature-toxicity relationships to yield predictive models. In this work, we adopted a principled approach to this problem by formulating a novel feature space based on intrinsic and extrinsic physicochemical properties, including periodic table properties but exclusive of in vitro characteristics such as cell line, cell type, and assay method. An optimal hypothesis space was developed by applying variance inflation analysis to the correlation structure of the features. Consequent to a stratified train-test split, the training dataset was balanced for the toxic outcomes and a mapping was then achieved from the normalized feature space to the toxicity class using various hyperparameter-tuned machine learning models, namely, logistic regression, random forest, support vector machines, and neural networks. Evaluation on an unseen test set yielded >96% balanced accuracy for the random forest, and neural network with one-hidden-layer models. The obtained cytotoxicity models are parsimonious, with intelligible inputs, and an embedded applicability check. Interpretability investigations of the models identified the key predictor variables of metal-oxide nanoparticle cytotoxicity. Our models could be applied on new, untested oxides, using a majority-voting ensemble classifier, NanoTox, that incorporates the best of the above models. NanoTox is the first open-source nanotoxicology pipeline, freely available under the GNU General Public License (https://github.com/NanoTox).

Finding Nano: Challenges Involved in Monitoring the Presence and Fate of Engineered Titanium Dioxide Nanoparticles in Aquatic Environments

In recent years, titanium dioxide (TiO2) has increasingly been used as an inorganic ultraviolet (UV) filter for sun protection. However, nano-TiO2 may also pose risks to the health of humans and the environment. Thus, to adequately assess its potential adverse effects, a comprehensive understanding of the behaviour and fate of TiO2 in different environments is crucial. Advances in analytical and modelling methods continue to improve researchers’ ability to quantify and determine the state of nano-TiO2 in various environments. However, due to the complexity of environmental and nanoparticle factors and their interplay, this remains a challenging and poorly resolved feat. This paper aims to provide a focused summary of key particle and environmental characteristics that influence the behaviour and fate of sunscreen-derived TiO2 in swimming pool water and natural aquatic environments and to review the current state-of-the-art of single particle inductively coupled plasma mass spectrometry (SP-ICP-MS) approaches to detect and characterise TiO2 nanoparticles in aqueous media. Furthermore, it critically analyses the capability of existing fate and transport models to predict environmental TiO2 levels. Four particle and environmental key factors that govern the fate and behaviour of TiO2 in aqueous environments are identified. A comparison of SP-ICP-MS studies reveals that it remains challenging to detect and characterise engineered TiO2 nanoparticles in various matrices and highlights the need for the development of new SP-ICP-MS pre-treatment and analysis approaches. This review shows that modelling studies are an essential addition to experimental studies, but they still lack in spatial and temporal resolution and mostly exclude surface transformation processes. Finally, this study identifies the use of Bayesian Network-based models as an underexplored but promising modelling tool to overcome data uncertainties and incorporates interconnected variables.

Probabilistic model for assessing occupational risk during the handling of nanomaterials

Exposure to nanomaterials (NMs) can be considered as human, occupational or environmental. Occupational exposure may be experienced by the workers and/or researchers who develop and produce these products and the hazards inherent to exposure are not yet fully known. Quantitative and qualitative methods are available to estimate the occupational risks associated with the handling of NMs, however, both have limitations. In this context, the objective of this study was to create a Bayesian network (BN) that will allow an assessment of the occupational risk associated with the handling of NMs in research laboratories. The BN was developed considered variables related to exposure, the hazards associated with NMs and also the existing control measures in the work environment, such as collective protection equipment (CPE), administrative measures and personal protection equipment (PPE). In addition to assessing the occupational risk, simulations were carried out by the laboratory manager to obtain information on which actions should be taken to reduce the risk. The development of a BN to assess the occupational risk associated with the handling of NMs is a novel aspect of this study. As a distinctive feature, the BN has measurement control variables in addition to considering CPE, administrative measures and PPE. An advantage of this network in relation to other risk assessment models is that it allows the easy execution of simulations and provides a guide for a decision making by identifying which actions should be taken to minimize the risk.

Artificial Intelligence and Machine Learning in Computational Nanotoxicology: Unlocking and Empowering Nanomedicine

Advances in nanomedicine, coupled with novel methods of creating advanced materials at the nanoscale, have opened new perspectives for the development of healthcare and medical products. Special attention must be paid toward safe design approaches for nanomaterial-based products. Recently, artificial intelligence (AI) and machine learning (ML) gifted the computational tool for enhancing and improving the simulation and modeling process for nanotoxicology and nanotherapeutics. In particular, the correlation of in vitro generated pharmacokinetics and pharmacodynamics to in vivo application scenarios is an important step toward the development of safe nanomedicinal products. This review portrays how in vitro and in vivo datasets are used in in silico models to unlock and empower nanomedicine. Physiologically based pharmacokinetic (PBPK) modeling and absorption, distribution, metabolism, and excretion (ADME)-based in silico methods along with dosimetry models as a focus area for nanomedicine are mainly described. The computational OMICS, colloidal particle determination, and algorithms to establish dosimetry for inhalation toxicology, and quantitative structure-activity relationships at nanoscale (nano-QSAR) are revisited. The challenges and opportunities facing the blind spots in nanotoxicology in this computationally dominated era are highlighted as the future to accelerate nanomedicine clinical translation.

Predicting In Vitro Neurotoxicity Induced by Nanoparticles Using Machine Learning

The practice of non-testing approaches in nanoparticles hazard assessment is necessary to identify and classify potential risks in a cost effective and timely manner. Machine learning techniques have been applied in the field of nanotoxicology with encouraging results. A neurotoxicity classification model for diverse nanoparticles is presented in this study. A data set created from multiple literature sources consisting of nanoparticles physicochemical properties, exposure conditions and in vitro characteristics is compiled to predict cell viability. Pre-processing techniques were applied such as normalization methods and two supervised instance methods, a synthetic minority over-sampling technique to address biased predictions and production of subsamples via bootstrapping. The classification model was developed using random forest and goodness-of-fit with additional robustness and predictability metrics were used to evaluate the performance. Information gain analysis identified the exposure dose and duration, toxicological assay, cell type, and zeta potential as the five most important attributes to predict neurotoxicity in vitro. This is the first tissue-specific machine learning tool for neurotoxicity prediction caused by nanoparticles in in vitro systems. The model performs better than non-tissue specific models.

QSAR without borders

Prediction of chemical bioactivity and physical properties has been one of the most important applications of statistical and more recently, machine learning and artificial intelligence methods in chemical sciences. This field of research, broadly known as quantitative structure-activity relationships (QSAR) modeling, has developed many important algorithms and has found a broad range of applications in physical organic and medicinal chemistry in the past 55+ years. This Perspective summarizes recent technological advances in QSAR modeling but it also highlights the applicability of algorithms, modeling methods, and validation practices developed in QSAR to a wide range of research areas outside of traditional QSAR boundaries including synthesis planning, nanotechnology, materials science, biomaterials, and clinical informatics. As modern research methods generate rapidly increasing amounts of data, the knowledge of robust data-driven modelling methods professed within the QSAR field can become essential for scientists working both within and outside of chemical research. We hope that this contribution highlighting the generalizable components of QSAR modeling will serve to address this challenge.

Nanotoxicology data for in silico tools: a literature review

The exercise of non-testing approaches in nanoparticles (NPs) hazard assessment is necessary for the risk assessment, considering cost and time efficiency, to identify, assess, and classify potential risks. One strategy for investigating the toxicological properties of a variety of NPs is by means of computational tools that decode how nano-specific features relate to toxicity and enable its prediction. This literature review records systematically the data used in published studies that predict nano (eco)-toxicological endpoints using machine learning models. Instead of seeking mechanistic interpretations this review maps the pathways followed, involving biological features in relation to NPs exposure, their physico-chemical characteristics and the most commonly predicted outcomes. The results, derived from published research of the last decade, are summarized visually, providing prior-based data mining paradigms to be readily used by the nanotoxicology community in computational studies.

Practices and Trends of Machine Learning Application in Nanotoxicology

Machine Learning (ML) techniques have been applied in the field of nanotoxicology with very encouraging results. Adverse effects of nanoforms are affected by multiple features described by theoretical descriptors, nano-specific measured properties, and experimental conditions. ML has been proven very helpful in this field in order to gain an insight into features effecting toxicity, predicting possible adverse effects as part of proactive risk analysis, and informing safe design. At this juncture, it is important to document and categorize the work that has been carried out. This study investigates and bookmarks ML methodologies used to predict nano (eco)-toxicological outcomes in nanotoxicology during the last decade. It provides a review of the sequenced steps involved in implementing an ML model, from data pre-processing, to model implementation, model validation, and applicability domain. The review gathers and presents the step-wise information on techniques and procedures of existing models that can be used readily to assemble new nanotoxicological in silico studies and accelerates the regulation of in silico tools in nanotoxicology. ML applications in nanotoxicology comprise an active and diverse collection of ongoing efforts, although it is still in their early steps toward a scientific accord, subsequent guidelines, and regulation adoption. This study is an important bookend to a decade of ML applications to nanotoxicology and serves as a useful guide to further in silico applications.

Bayesian Network Resource for Meta-Analysis: Cellular Toxicity of Quantum Dots

A web-based resource for meta-analysis of nanomaterials toxicity is developed whereby the utility of Bayesian networks (BNs) is illustrated for exploring the cellular toxicity of Cd-containing quantum dots (QDs). BN models are developed based on a dataset compiled from 517 publications comprising 3028 cell viability data samples and 837 IC₅₀ values. BN QD toxicity (BN-QDTox) models are developed using both continuous (i.e., numerical) and categorical attributes. Using these models, the most relevant attributes identified for correlating IC₅₀ are: QD diameter, exposure time, surface ligand, shell, assay type, surface modification, and surface charge, with the addition of QD concentration for the cell viability analysis. Data exploration via BN models further enables identification of possible association rules for QDs cellular toxicity. The BN models as web-based applications can be used for rapid intelligent query of the available body of evidence for a given nanomaterial and can be readily updated as the body of knowledge expands.

Application of Bayesian networks in determining nanoparticle-induced cellular outcomes using transcriptomics

Inroads have been made in our understanding of the risks posed to human health and the environment by nanoparticles (NPs) but this area requires continuous research and monitoring. Machine learning techniques have been applied to nanotoxicology with very encouraging results. This study deals with bridging physicochemical properties of NPs, experimental exposure conditions and in vitro characteristics with biological effects of NPs on a molecular cellular level from transcriptomics studies. The bridging is done by developing and implementing Bayesian Networks (BNs) with or without data preprocessing. The BN structures are derived either automatically or methodologically and compared. Early stage nanotoxicity measurements represent a challenge, not least when attempting to predict adverse outcomes and modeling is critical to understanding the biological effects of exposure to NPs. The preprocessed data-driven BN showed improved performance over automatically structured BN and the BN with unprocessed datasets. The prestructured BN captures inter relationships between NP properties, exposure condition and in vitro characteristics and links those with cellular effects based on statistic correlation findings. Information gain analysis showed that exposure dose, NP and cell line variables were the most influential attributes in predicting the biological effects. The BN methodology proposed in this study successfully predicts a number of toxicologically relevant cellular disrupted biological processes such as cell cycle and proliferation pathways, cell adhesion and extracellular matrix responses, DNA damage and repair mechanisms etc., with a success rate >80%. The model validation from independent data shows a robust and promising methodology for incorporating transcriptomics outcomes in a hazard and, by extension, risk assessment modeling framework by predicting affected cellular functions from experimental conditions.

Exploring the importance of mixed autogenous vaccines as a potential determinant of lung consolidation in lambs using Bayesian networks

Bayesian networks are used to evaluate the effectiveness of mixed autogenous vaccines in fattening lambs to prevent the ovine respiratory syndrome. An experiment was performed with 460 fattening lambs, which were clustered into four groups according to the kind of vaccine received (Pasteurella spp., Mycoplasma spp., Mixed Mycoplasma-Pasteurella or placebo). After slaughtering, lungs were collected, and macroscopic and microscopic studies were performed. A microbiological study was carried out to evaluate the presence of Mycoplasma spp. and Pasteurellaceae by conventional culture and identification by nested polymerase chain reaction. To the best of the authors' knowledge, Bayesian networks have not been used to evaluate the effect of vaccines on the absence/presence of lung consolidation. Our results revealed that the use of mixed autogenous vaccines can decrease lung consolidation from 15.75% (12.42-19.08) to 9.24% (6.59-11.89). Therefore, the use of these autogenous vaccines in farms could be considered an effective control tool against ovine respiratory syndrome.

Machine learning prediction of nanoparticle in vitro toxicity: A comparative study of classifiers and ensemble-classifiers using the Copeland Index

Nano-Particles (NPs) are well established as important components across a broad range of products from cosmetics to electronics. Their utilization is increasing with their significant economic and societal potential yet to be fully realized. Inroads have been made in our understanding of the risks posed to human health and the environment by NPs but this area will require continuous research and monitoring. In recent years Machine Learning (ML) techniques have exploited large datasets and computation power to create breakthroughs in diverse fields from facial recognition to genomics. More recently, ML techniques have been applied to nanotoxicology with very encouraging results. In this study, categories of ML classifiers (rules, trees, lazy, functions and bayes) were compared using datasets from the Safe and Sustainable Nanotechnology (S2NANO) database to investigate their performance in predicting NPs in vitro toxicity. Physicochemical properties, toxicological and quantum-mechanical attributes and in vitro experimental conditions were used as input variables to predict the toxicity of NPs based on cell viability. Voting, an ensemble meta-classifier, was used to combine base models to optimize the classification prediction of toxicity. To facilitate inter-comparison, a Copeland Index was applied that ranks the classifiers according to their performance and suggested the optimal classifier. Neural Network (NN) and Random forest (RF) showed the best performance in the majority of the datasets used in this study. However, the combination of classifiers demonstrated an improved prediction resulting meta-classifier to have higher indices. This proposed Copeland Index can now be used by researchers to identify and clearly prioritize classifiers in order to achieve more accurate classification predictions for NP toxicity for a given dataset.

Hazard Screening Methods for Nanomaterials: A Comparative Study

Hazard identification is the key step in risk assessment and management of manufactured nanomaterials (NM). However, the rapid commercialisation of nano-enabled products continues to out-pace the development of a prudent risk management mechanism that is widely accepted by the scientific community and enforced by regulators. However, a growing body of academic literature is developing promising quantitative methods. Two approaches have gained significant currency. Bayesian networks (BN) are a probabilistic, machine learning approach while the weight of evidence (WoE) statistical framework is based on expert elicitation. This comparative study investigates the efficacy of quantitative WoE and Bayesian methodologies in ranking the potential hazard of metal and metal-oxide NMs—TiO₂, Ag, and ZnO. This research finds that hazard ranking is consistent for both risk assessment approaches. The BN and WoE models both utilize physico-chemical, toxicological, and study type data to infer the hazard potential. The BN exhibits more stability when the models are perturbed with new data. The BN has the significant advantage of self-learning with new data; however, this assumes all input data is equally valid. This research finds that a combination of WoE that would rank input data along with the BN is the optimal hazard assessment framework.