CATMoS: Collaborative Acute Toxicity Modeling Suite

Where developmental toxicity meets explainable artificial intelligence: state-of-the-art and perspectives

INTRODUCTION

The application of Artificial Intelligence (AI) to predictive toxicology is rapidly increasing, particularly aiming to develop non-testing methods that effectively address ethical concerns and reduce economic costs. In this context, Developmental Toxicity (Dev Tox) stands as a key human health endpoint, especially significant for safeguarding maternal and child well-being.

AREAS COVERED

This review outlines the existing methods employed in Dev Tox predictions and underscores the benefits of utilizing New Approach Methodologies (NAMs), specifically focusing on eXplainable Artificial Intelligence (XAI), which proves highly efficient in constructing reliable and transparent models aligned with recommendations from international regulatory bodies.

EXPERT OPINION

The limited availability of high-quality data and the absence of dependable Dev Tox methodologies render XAI an appealing avenue for systematically developing interpretable and transparent models, which hold immense potential for both scientific evaluations and regulatory decision-making.

The rat acute oral toxicity of trifluoromethyl compounds (TFMs): a computational toxicology study combining the 2D-QSTR, read-across and consensus modeling methods

All areas of the modern society are affected by fluorine chemistry. In particular, fluorine plays an important role in medical, pharmaceutical and agrochemical sciences. Amongst various fluoro-organic compounds, trifluoromethyl (CF3) group is valuable in applications such as pharmaceuticals, agrochemicals and industrial chemicals. In the present study, following the strict OECD modelling principles, a quantitative structure-toxicity relationship (QSTR) modelling for the rat acute oral toxicity of trifluoromethyl compounds (TFMs) was established by genetic algorithm-multiple linear regression (GA-MLR) approach. All developed models were evaluated by various state-of-the-art validation metrics and the OECD principles. The best QSTR model included nine easily interpretable 2D molecular descriptors with clear physical and chemical significance. The mechanistic interpretation showed that the atom-type electro-topological state indices, molecular connectivity, ionization potential, lipophilicity and some autocorrelation coefficients are the main factors contributing to the acute oral toxicity of TFMs against rats. To validate that the selected 2D descriptors can effectively characterize the toxicity, we performed the chemical read-across analysis. We also compared the best QSTR model with public OPERA tool to demonstrate the reliability of the predictions. To further improve the prediction range of the QSTR model, we performed the consensus modelling. Finally, the optimum QSTR model was utilized to predict a true external set containing many untested/unknown TFMs for the first time. Overall, the developed model contributes to a more comprehensive safety assessment approach for novel CF3-containing pharmaceuticals or chemicals, reducing unnecessary chemical synthesis whilst saving the development cost of new drugs.

Applying new approach methodologies to assess next-generation tobacco and nicotine products

In vitro toxicology research has accelerated with the use of in silico, computational approaches and human in vitro tissue systems, facilitating major improvements evaluating the safety and health risks of novel consumer products. Innovation in molecular and cellular biology has shifted testing paradigms, with less reliance on low-throughput animal data and greater use of medium- and high-throughput in vitro cellular screening approaches. These new approach methodologies (NAMs) are being implemented in other industry sectors for chemical testing, screening candidate drugs and prototype consumer products, driven by the need for reliable, human-relevant approaches. Routine toxicological methods are largely unchanged since development over 50 years ago, using high-doses and often employing in vivo testing. Several disadvantages are encountered conducting or extrapolating data from animal studies due to differences in metabolism or exposure. The last decade saw considerable advancement in the development of in vitro tools and capabilities, and the challenges of the next decade will be integrating these platforms into applied product testing and acceptance by regulatory bodies. Governmental and validation agencies have launched and applied frameworks and "roadmaps" to support agile validation and acceptance of NAMs. Next-generation tobacco and nicotine products (NGPs) have the potential to offer reduced risks to smokers compared to cigarettes. These include heated tobacco products (HTPs) that heat but do not burn tobacco; vapor products also termed electronic nicotine delivery systems (ENDS), that heat an e-liquid to produce an inhalable aerosol; oral smokeless tobacco products (e.g., Swedish-style snus) and tobacco-free oral nicotine pouches. With the increased availability of NGPs and the requirement of scientific studies to support regulatory approval, NAMs approaches can supplement the assessment of NGPs. This review explores how NAMs can be applied to assess NGPs, highlighting key considerations, including the use of appropriate in vitro model systems, deploying screening approaches for hazard identification, and the importance of test article characterization. The importance and opportunity for fit-for-purpose testing and method standardization are discussed, highlighting the value of industry and cross-industry collaborations. Supporting the development of methods that are accepted by regulatory bodies could lead to the implementation of NAMs for tobacco and nicotine NGP testing.

Cancer drug sensitivity estimation using modular deep Graph Neural Networks

Computational drug sensitivity models have the potential to improve therapeutic outcomes by identifying targeted drugs components that are tailored to the transcriptomic profile of a given primary tumor. The SMILES representation of molecules that is used by state-of-the-art drug-sensitivity models is not conducive for neural networks to generalize to new drugs, in part because the distance between atoms does not generally correspond to the distance between their representation in the SMILES strings. Graph-attention networks, on the other hand, are high-capacity models that require large training-data volumes which are not available for drug-sensitivity estimation. We develop a modular drug-sensitivity graph-attentional neural network. The modular architecture allows us to separately pre-train the graph encoder and graph-attentional pooling layer on related tasks for which more data are available. We observe that this model outperforms reference models for the use cases of precision oncology and drug discovery; in particular, it is better able to predict the specific interaction between drug and cell line that is not explained by the general cytotoxicity of the drug and the overall survivability of the cell line. The complete source code is available at https://zenodo.org/doi/10.5281/zenodo.8020945. All experiments are based on the publicly available GDSC data.

The estimation of acute oral toxicity (LD50) of G-series organophosphorus-based chemical warfare agents using quantitative and qualitative toxicology in silico methods

The idea of this study was the estimation of the theoretical acute toxicity (t-LD50, rat, oral dose) of organophosphorus-based chemical warfare agents from the G-series (n = 12) using different in silico methods. Initially identified in Germany, the G-type nerve agents include potent compounds such as tabun, sarin, and soman. Despite their historical significance, there is a noticeable gap in acute toxicity data for these agents. This study employs qualitative (STopTox and AdmetSAR) and quantitative (TEST; CATMoS; ProTox-II and QSAR Toolbox) in silico methods to predict LD50 values, offering an ethical alternative to animal testing. Additionally, we conducted quantitative extrapolation from animals, and the results of qualitative tests confirmed the acute toxicity potential of these substances and enabled the identification of toxicophoric groups. According to our estimations, the most lethal agents within this category were GV, soman (GD), sarin (GB), thiosarin (GBS), and chlorosarin (GC), with t-LD50 values (oral administration, extrapolated from rat to human) of 0.05 mg/kg bw, 0.08 mg/kg bw, 0.12 mg/kg bw, 0.15 mg/kg bw, and 0.17 mg/kg bw, respectively. On the contrary, compounds with a cycloalkane attached to the phospho-oxygen linkage, specifically methyl cyclosarin and cyclosarin, were found to be the least toxic, with values of 2.28 mg/kg bw and 3.03 mg/kg bw. The findings aim to fill the knowledge gap regarding the acute toxicity of these agents, highlighting the need for modern toxicological methods that align with ethical considerations, next-generation risk assessment (NGRA) and the 3Rs (replacement, reduction and refinement) principles.

From Pipeline to Plant Protection Products: Using New Approach Methodologies (NAMs) in Agrochemical Safety Assessment

The human population will be approximately 9.7 billion by 2050, and food security has been identified as one of the key issues facing the global population. Agrochemicals are an important tool available to farmers that enable high crop yields and continued access to healthy foods, but the average new agrochemical active ingredient takes more than ten years, 350 million dollars, and 20,000 animals to develop and register. The time, monetary, and animal costs incentivize the use of New Approach Methodologies (NAMs) in early-stage screening to prioritize chemical candidates. This review outlines NAMs that are currently available or can be adapted for use in early-stage screening agrochemical programs. It covers new in vitro screens that are on the horizon in key areas of regulatory concern. Overall, early-stage screening with NAMs enables the prioritization of development for agrochemicals without human and environmental health concerns through a more directed, agile, and iterative development program before animal-based regulatory testing is even considered.

Probabilistic Reference and 10% Effect Concentrations for Characterizing Inhalation Non-cancer and Developmental/Reproductive Effects for 2,160 Substances

Chemicals assessment and management frameworks rely on regulatory toxicity values, which are based on points of departure (POD) identified following rigorous dose-response assessments. Yet, regulatory PODs and toxicity values for inhalation exposure (i.e., reference concentrations [RfCs]) are available for only ∼200 chemicals. To address this gap, we applied a workflow to determine surrogate inhalation route PODs and corresponding toxicity values, where regulatory assessments are lacking. We curated and selected inhalation in vivo data from the U.S. EPA's ToxValDB and adjusted reported effect values to chronic human equivalent benchmark concentrations (BMCh) following the WHO/IPCS framework. Using ToxValDB chemicals with existing PODs associated with regulatory toxicity values, we found that the 25th %-ile of a chemical's BMCh distribution (POD p 25 BMC h ) could serve as a suitable surrogate for regulatory PODs (Q2 ≥ 0.76, RSE ≤ 0.82 log10 units). We applied this approach to derive POD p 25 BMC h for 2,095 substances with general non-cancer toxicity effects and 638 substances with reproductive/developmental toxicity effects, yielding a total coverage of 2,160 substances. From these POD p 25 BMC h , we derived probabilistic RfCs and human population effect concentrations. With this work, we have expanded the number of chemicals with toxicity values available, thereby enabling a much broader coverage for inhalation risk and impact assessment.

Two-Stage Machine Learning-Based Approach to Predict Points of Departure for Human Noncancer and Developmental/Reproductive Effects

Chemical points of departure (PODs) for critical health effects are crucial for evaluating and managing human health risks and impacts from exposure. However, PODs are unavailable for most chemicals in commerce due to a lack of in vivo toxicity data. We therefore developed a two-stage machine learning (ML) framework to predict human-equivalent PODs for oral exposure to organic chemicals based on chemical structure. Utilizing ML-based predictions for structural/physical/chemical/toxicological properties from OPERA 2.9 as features (Stage 1), ML models using random forest regression were trained with human-equivalent PODs derived from in vivo data sets for general noncancer effects (n = 1,791) and reproductive/developmental effects (n = 2,228), with robust cross-validation for feature selection and estimating generalization errors (Stage 2). These two-stage models accurately predicted PODs for both effect categories with cross-validation-based root-mean-squared errors less than an order of magnitude. We then applied one or both models to 34,046 chemicals expected to be in the environment, revealing several thousand chemicals of moderate concern and several hundred chemicals of high concern for health effects at estimated median population exposure levels. Further application can expand by orders of magnitude the coverage of organic chemicals that can be evaluated for their human health risks and impacts.

Evaluation of in silico model predictions for mammalian acute oral toxicity and regulatory application in pesticide hazard and risk assessment

The United States Environmental Protection Agency (USEPA) uses the lethal dose 50% (LD50) value from in vivo rat acute oral toxicity studies for pesticide product label precautionary statements and environmental risk assessment (RA). The Collaborative Acute Toxicity Modeling Suite (CATMoS) is a quantitative structure-activity relationship (QSAR)-based in silico approach to predict rat acute oral toxicity that has the potential to reduce animal use when registering a new pesticide technical grade active ingredient (TGAI). This analysis compared LD50 values predicted by CATMoS to empirical values from in vivo studies for the TGAIs of 177 conventional pesticides. The accuracy and reliability of the model predictions were assessed relative to the empirical data in terms of USEPA acute oral toxicity categories and discrete LD50 values for each chemical. CATMoS was most reliable at placing pesticide TGAIs in acute toxicity categories III (>500-5000 mg/kg) and IV (>5000 mg/kg), with 88% categorical concordance for 165 chemicals with empirical in vivo LD50 values ≥ 500 mg/kg. When considering an LD50 for RA, CATMoS predictions of 2000 mg/kg and higher were found to agree with empirical values from limit tests (i.e., single, high-dose tests) or definitive results over 2000 mg/kg with few exceptions.

The acute toxicity of Novichok's degradation products using quantitative and qualitative toxicology in silico methods

The emergence of Novichok agents, potent organophosphorus nerve agents, has spurred the demand for advanced analytical methods and toxicity assessments as a result of their involvement in high-profile incidents. This study focuses on the degradation products of Novichok agents, particularly their potential toxic effects on biological systems. Traditional in vivo methods for toxicity evaluation face ethical and practical constraints, prompting a shift toward in silico toxicology research. In this context, we conducted a comprehensive qualitative and quantitative analysis of acute oral toxicity (AOT) for Novichok degradation products, using various in silico methods, including TEST, CATMoS, ProTox-II, ADMETlab, ACD/Labs Percepta, and QSAR Toolbox. Adopting these methodologies aligns with the 3Rs principle, emphasising Replacement, Reduction, and Refinement in the realm of toxicological studies. Qualitative assessments with STopTox and admetSAR revealed toxic profiles for all degradation products, with predicted toxicophores highlighting structural features responsible for toxicity. Quantitative predictions yielded varied estimates of acute oral toxicity, with the most toxic degradation products being EOPAA, MOPGA, MOPAA, MPGA, EOPGA, and MPAA, respectively. Structural modifications common to all examined hydrolytic degradation products involve substituting the fluorine atom with a hydroxyl group, imparting consequential effects on toxicity. The need for sophisticated analytical techniques for identifying and quantifying Novichok degradation products is underscored due to their inherent reactivity. This study represents a crucial step in unravelling the complexities of Novichok toxicity, highlighting the ongoing need for research into its degradation processes to refine analytical methodologies and fortify readiness against potential threats.

Toxicological effects assessment for wildlife in the 21st century: Review of current methods and recommendations for a path forward

Model species (e.g., granivorous gamebirds, waterfowl, passerines, domesticated rodents) have been used for decades in guideline laboratory tests to generate survival, growth, and reproductive data for prospective ecological risk assessments (ERAs) for birds and mammals, while officially adopted risk assessment schemes for amphibians and reptiles do not exist. There are recognized shortcomings of current in vivo methods as well as uncertainty around the extent to which species with different life histories (e.g., terrestrial amphibians, reptiles, bats) than these commonly used models are protected by existing ERA frameworks. Approaches other than validating additional animal models for testing are being developed, but the incorporation of such new approach methodologies (NAMs) into risk assessment frameworks will require robust validations against in vivo responses. This takes time, and the ability to extrapolate findings from nonanimal studies to organism- and population-level effects in terrestrial wildlife remains weak. Failure to adequately anticipate and predict hazards could have economic and potentially even legal consequences for regulators and product registrants. In order to be able to use fewer animals or replace them altogether in the long term, vertebrate use and whole organism data will be needed to provide data for NAM validation in the short term. Therefore, it is worth investing resources for potential updates to existing standard test guidelines used in the laboratory as well as addressing the need for clear guidance on the conduct of field studies. Herein, we review the potential for improving standard in vivo test methods and for advancing the use of field studies in wildlife risk assessment, as these tools will be needed in the foreseeable future. Integr Environ Assess Manag 2024;20:699-724. © 2023 His Majesty the King in Right of Canada and The Authors. Integrated Environmental Assessment and Management published by Wiley Periodicals LLC on behalf of Society of Environmental Toxicology & Chemistry (SETAC). Reproduced with the permission of the Minister of Environment and Climate Change Canada. This article has been contributed to by U.S. Government employees and their work is in the public domain in the USA.

In Silico Prediction of Oral Acute Rodent Toxicity Using Consensus Machine Learning

Acute oral toxicity (AOT) is required for the classification and labeling of chemicals according to the global harmonized system (GHS). Acute oral toxicity studies are optimized to minimize the use of animals. However, with the advent of the three Rs principles and machine learning in toxicology, alternative in silico methods became a reasonable alternative approach for addressing the AOT of new chemical matter. Here, we describe the compilation of AOT data from a commercial database and the development of a consensus classification model after evaluating different combinations of molecular representations and machine learning algorithms. The model shows significantly better performance compared to publicly available AOT models. Its performance was evaluated on an external validation data set, which was compiled from the literature, and an applicability domain was deduced.

Expanding Predictive Capacities in Toxicology: Insights from Hackathon-Enhanced Data and Model Aggregation

In the realm of predictive toxicology for small molecules, the applicability domain of QSAR models is often limited by the coverage of the chemical space in the training set. Consequently, classical models fail to provide reliable predictions for wide classes of molecules. However, the emergence of innovative data collection methods such as intensive hackathons have promise to quickly expand the available chemical space for model construction. Combined with algorithmic refinement methods, these tools can address the challenges of toxicity prediction, enhancing both the robustness and applicability of the corresponding models. This study aimed to investigate the roles of gradient boosting and strategic data aggregation in enhancing the predictivity ability of models for the toxicity of small organic molecules. We focused on evaluating the impact of incorporating fragment features and expanding the chemical space, facilitated by a comprehensive dataset procured in an open hackathon. We used gradient boosting techniques, accounting for critical features such as the structural fragments or functional groups often associated with manifestations of toxicity.

Systematic approaches to machine learning models for predicting pesticide toxicity

Pesticides play an important role in modern agriculture by protecting crops from pests and diseases. However, the negative consequences of pesticides, such as environmental contamination and adverse effects on human and ecological health, underscore the importance of accurate toxicity predictions. To address this issue, artificial intelligence models have emerged as valuable methods for predicting the toxicity of organic compounds. In this review article, we explore the application of machine learning (ML) for pesticide toxicity prediction. This review provides a detailed summary of recent developments, prediction models, and datasets used for pesticide toxicity prediction. In this analysis, we compared the results of several algorithms that predict the harmfulness of various classes of pesticides. Furthermore, this review article identified emerging trends and areas for future direction, showcasing the transformative potential of machine learning in promoting safer pesticide usage and sustainable agriculture.

Deus Ex Machina? The Rise of Artificial Intelligence in Toxicology

Artificial intelligence (AI) is rising rapidly, driven by big data, complex algorithms, and computing resources. Current research presented at the American Chemical Society Fall 2023 Meeting demonstrates AI to be a valuable predictive and supporting tool across all facets of toxicology.

VenomPred 2.0: A Novel In Silico Platform for an Extended and Human Interpretable Toxicological Profiling of Small Molecules

The application of artificial intelligence and machine learning (ML) methods is becoming increasingly popular in computational toxicology and drug design; it is considered as a promising solution for assessing the safety profile of compounds, particularly in lead optimization and ADMET studies, and to meet the principles of the 3Rs, which calls for the replacement, reduction, and refinement of animal testing. In this context, we herein present the development of VenomPred 2.0 (http://www.mmvsl.it/wp/venompred2/), the new and improved version of our free of charge web tool for toxicological predictions, which now represents a powerful web-based platform for multifaceted and human-interpretable in silico toxicity profiling of chemicals. VenomPred 2.0 presents an extended set of toxicity endpoints (androgenicity, skin irritation, eye irritation, and acute oral toxicity, in addition to the already available carcinogenicity, mutagenicity, hepatotoxicity, and estrogenicity) that can be evaluated through an exhaustive consensus prediction strategy based on multiple ML models. Moreover, we also implemented a new utility based on the Shapley Additive exPlanations (SHAP) method that allows human interpretable toxicological profiling of small molecules, highlighting the features that strongly contribute to the toxicological predictions in order to derive structural toxicophores.

Artificial intelligence (AI)-it's the end of the tox as we know it (and I feel fine)

The rapid progress of AI impacts diverse scientific disciplines, including toxicology, and has the potential to transform chemical safety evaluation. Toxicology has evolved from an empirical science focused on observing apical outcomes of chemical exposure, to a data-rich field ripe for AI integration. The volume, variety and velocity of toxicological data from legacy studies, literature, high-throughput assays, sensor technologies and omics approaches create opportunities but also complexities that AI can help address. In particular, machine learning is well suited to handle and integrate large, heterogeneous datasets that are both structured and unstructured-a key challenge in modern toxicology. AI methods like deep neural networks, large language models, and natural language processing have successfully predicted toxicity endpoints, analyzed high-throughput data, extracted facts from literature, and generated synthetic data. Beyond automating data capture, analysis, and prediction, AI techniques show promise for accelerating quantitative risk assessment by providing probabilistic outputs to capture uncertainties. AI also enables explanation methods to unravel mechanisms and increase trust in modeled predictions. However, issues like model interpretability, data biases, and transparency currently limit regulatory endorsement of AI. Multidisciplinary collaboration is needed to ensure development of interpretable, robust, and human-centered AI systems. Rather than just automating human tasks at scale, transformative AI can catalyze innovation in how evidence is gathered, data are generated, hypotheses are formed and tested, and tasks are performed to usher new paradigms in chemical safety assessment. Used judiciously, AI has immense potential to advance toxicology into a more predictive, mechanism-based, and evidence-integrated scientific discipline to better safeguard human and environmental wellbeing across diverse populations.

Free and open-source QSAR-ready workflow for automated standardization of chemical structures in support of QSAR modeling

The rapid increase of publicly available chemical structures and associated experimental data presents a valuable opportunity to build robust QSAR models for applications in different fields. However, the common concern is the quality of both the chemical structure information and associated experimental data. This is especially true when those data are collected from multiple sources as chemical substance mappings can contain many duplicate structures and molecular inconsistencies. Such issues can impact the resulting molecular descriptors and their mappings to experimental data and, subsequently, the quality of the derived models in terms of accuracy, repeatability, and reliability. Herein we describe the development of an automated workflow to standardize chemical structures according to a set of standard rules and generate two and/or three-dimensional "QSAR-ready" forms prior to the calculation of molecular descriptors. The workflow was designed in the KNIME workflow environment and consists of three high-level steps. First, a structure encoding is read, and then the resulting in-memory representation is cross-referenced with any existing identifiers for consistency. Finally, the structure is standardized using a series of operations including desalting, stripping of stereochemistry (for two-dimensional structures), standardization of tautomers and nitro groups, valence correction, neutralization when possible, and then removal of duplicates. This workflow was initially developed to support collaborative modeling QSAR projects to ensure consistency of the results from the different participants. It was then updated and generalized for other modeling applications. This included modification of the "QSAR-ready" workflow to generate "MS-ready structures" to support the generation of substance mappings and searches for software applications related to non-targeted analysis mass spectrometry. Both QSAR and MS-ready workflows are freely available in KNIME, via standalone versions on GitHub, and as docker container resources for the scientific community. Scientific contribution: This work pioneers an automated workflow in KNIME, systematically standardizing chemical structures to ensure their readiness for QSAR modeling and broader scientific applications. By addressing data quality concerns through desalting, stereochemistry stripping, and normalization, it optimizes molecular descriptors' accuracy and reliability. The freely available resources in KNIME, GitHub, and docker containers democratize access, benefiting collaborative research and advancing diverse modeling endeavors in chemistry and mass spectrometry.

Development and application of consensus in silico models for advancing high-throughput toxicological predictions

Computational toxicology models have been successfully implemented to prioritize and screen chemicals. There are numerous in silico (quantitative) structure-activity relationship ([Q]SAR) models for the prediction of a range of human-relevant toxicological endpoints, but for a given endpoint and chemical, not all predictions are identical due to differences in their training sets, algorithms, and methodology. This poses an issue for high-throughput screening of a large chemical inventory as it necessitates several models to cover diverse chemistries but will then generate data conflicts. To address this challenge, we developed a consensus modeling strategy to combine predictions obtained from different existing in silico (Q)SAR models into a single predictive value while also expanding chemical space coverage. This study developed consensus models for nine toxicological endpoints relating to estrogen receptor (ER) and androgen receptor (AR) interactions (i.e., binding, agonism, and antagonism) and genotoxicity (i.e., bacterial mutation, in vitro chromosomal aberration, and in vivo micronucleus). Consensus models were created by combining different (Q)SAR models using various weighting schemes. As a multi-objective optimization problem, there is no single best consensus model, and therefore, Pareto fronts were determined for each endpoint to identify the consensus models that optimize the multiple-criterion decisions simultaneously. Accordingly, this work presents sets of solutions for each endpoint that contain the optimal combination, regardless of the trade-off, with the results demonstrating that the consensus models improved both the predictive power and chemical space coverage. These solutions were further analyzed to find trends between the best consensus models and their components. Here, we demonstrate the development of a flexible and adaptable approach for in silico consensus modeling and its application across nine toxicological endpoints related to ER activity, AR activity, and genotoxicity. These consensus models are developed to be integrated into a larger multi-tier NAM-based framework to prioritize chemicals for further investigation and support the transition to a non-animal approach to risk assessment in Canada.

Analysis of health concerns not addressed by REACH for low tonnage chemicals and opportunities for new approach methodology

In Registration, Evaluation, Authorisation and Restriction of Chemicals (REACH) the criterion for deciding the studies that must be performed is the annual tonnage of the chemical manufactured or imported into the EU. The annual tonnage may be considered as a surrogate for levels of human exposure but this does not take into account the physico-chemical properties and use patterns that determine exposure. Chemicals are classified using data from REACH under areas of health concern covering effects on the skin and eye; sensitisation; acute, repeated and prolonged systemic exposure; effects on genetic material; carcinogenicity; and reproduction and development. We analysed the mandated study lists under REACH for each annual tonnage band in terms of the information they provide on each of the areas of health concern. Using the European Chemicals Agency (ECHA) REACH Registration data base of over 20,000 registered substances, we found that only 19% of registered substances have datasets on all areas of health concern. Information limited to acute exposure, sensitisation and genotoxicity was found for 62%. The analysis highlighted the shortfall of information mandated for substances in the lower tonnage bands. Deploying New Approach Methodologies (NAMs) at this lower tonnage band to assess health concerns which are currently not covered by REACH, such as repeat and extended exposure and carcinogenicity, would provide additional information and would be a way for registrants and regulators to gain experience in the use of NAMs. There are currently projects in Europe aiming to develop NAM-based assessment frameworks and they could find their first use in assessing low tonnage chemicals once confidence has been gained by their evaluation with data rich chemicals.

The application of acute oral toxicity computational models in dangerous goods classification

Acute oral toxicity (AOT) data inform the acute toxicity potential of a compound and guides occupational safety and transportation practices. AOT data enable the categorization of a chemical into the appropriate AOT Globally Harmonized System (GHS) category based on the severity of the hazard. AOT data are also utilized to identify compounds that are Dangerous Goods (DGs) and subsequent transportation guidance for shipping of these hazardous materials. Proper identification of DGs is challenging for novel compounds that lack data. It is not feasible to err on the side of caution for all compounds lacking AOT data and to designate them as DGs, as shipping a compound as a DG has cost, resource, and time implications. With the wealth of available historical AOT data, AOT testing approaches are evolving, and in silico AOT models are emerging as tools that can be utilized with confidence to assess the acute toxicity potential of de novo molecules. Such approaches align with the 3R principles, offering a reduction or even replacement of traditional in vivo testing methods and can also be leveraged for product stewardship purposes. Utilizing proprietary historical in vivo AOT data for 210 pharmaceutical compounds (PCs), we evaluated the performance of two established in silico AOT programs: the Leadscope AOT Model Suite and the Collaborative Acute Toxicity Modeling Suite. These models accurately identified 94% and 97% compounds that were not DGs (GHS categories 4, 5, and not classified (NC)) suggesting that the models are fit-for-purpose in identifying PCs with low acute oral toxicity potential (LD50 >300 mg/kg). Utilization of these models to identify compounds that are not DGs can enable them to be de-prioritized for in vivo testing. This manuscript provides a detailed evaluation and assessment of the two models and recommends the most suitable applications of such models.

The worm Adult Activity Test (wAAT): A de novo mathematical model for detecting acute chemical effects in Caenorhabditis elegans

We have adapted a semiautomated method for tracking Caenorhabditis elegans spontaneous locomotor activity into a quantifiable assay by developing a sophisticated method for analyzing the time course of measured activity. The 16-h worm Adult Activity Test (wAAT) can be used to measure C. elegans activity levels for efficient screening for pharmacological and toxicity-induced effects. As with any apical endpoint assay, the wAAT is mode of action agnostic, allowing for detection of effects from a broad spectrum of response pathways. With caffeine as a model mild stimulant, the wAAT showed transient hyperactivity followed by reversion to baseline. Mercury chloride (HgCl2 ) produced an early dose-response hyperactivity phase followed by pronounced hypoactivity, a behavior pattern we have termed a toxicant "escape response." Methylmercury chloride (meHgCl) produced a similar pattern to HgCl2 , but at much lower concentrations, a weaker hyperactivity response, and more pronounced hypoactivity. Sodium arsenite (NaAsO2 ) and dimethylarsinic acid (DMA) induced hypoactivity at high concentrations. Acute toxicity, as measured by hypoactivity in C. elegans adults, was ranked: meHgCl > HgCl2  > NaAsO2  = DMA. Caffeine was not toxic with the wAAT at tested concentrations. Methods for conducting the wAAT are described, along with instructions for preparing C. elegans Habitation Medium, a liquid nutrient medium that allows for developmental timing equivalent to that found with C. elegans grown on agar with OP50 Escherichia coli feeder cultures. A de novo mathematical parametric model for adult C. elegans activity and the application of this model in ranking exposure toxicity are presented.

Review of machine learning and deep learning models for toxicity prediction

The ever-increasing number of chemicals has raised public concerns due to their adverse effects on human health and the environment. To protect public health and the environment, it is critical to assess the toxicity of these chemicals. Traditional in vitro and in vivo toxicity assays are complicated, costly, and time-consuming and may face ethical issues. These constraints raise the need for alternative methods for assessing the toxicity of chemicals. Recently, due to the advancement of machine learning algorithms and the increase in computational power, many toxicity prediction models have been developed using various machine learning and deep learning algorithms such as support vector machine, random forest, k-nearest neighbors, ensemble learning, and deep neural network. This review summarizes the machine learning- and deep learning-based toxicity prediction models developed in recent years. Support vector machine and random forest are the most popular machine learning algorithms, and hepatotoxicity, cardiotoxicity, and carcinogenicity are the frequently modeled toxicity endpoints in predictive toxicology. It is known that datasets impact model performance. The quality of datasets used in the development of toxicity prediction models using machine learning and deep learning is vital to the performance of the developed models. The different toxicity assignments for the same chemicals among different datasets of the same type of toxicity have been observed, indicating benchmarking datasets is needed for developing reliable toxicity prediction models using machine learning and deep learning algorithms. This review provides insights into current machine learning models in predictive toxicology, which are expected to promote the development and application of toxicity prediction models in the future.

Predictive Models Based on Molecular Images and Molecular Descriptors for Drug Screening

Various toxicity and pharmacokinetic evaluations as screening experiments are needed at the drug discovery stage. Currently, to reduce the use of animal experiments and developmental expenses, the development of high-performance predictive models based on quantitative structure-activity relationship analysis is desired. From these evaluation targets, we selected 50% lethal dose (LD50), blood-brain barrier penetration (BBBP), and the clearance (CL) pathway for this investigation and constructed predictive models for each target using 636-11,886 compounds. First, we constructed predictive models using the DeepSnap-deep learning (DL) method and images of compounds as features. The calculated area under the curve (AUC) and balanced accuracy (BAC) were, respectively, 0.887 and 0.818 for LD50, 0.893 and 0.824 for BBBP, and 0.883 and 0.763 for the CL pathway. Next, molecular descriptors (MDs) of compounds were calculated using Molecular Operating Environment, alvaDesc, and ADMET Predictor to construct predictive models using the MD-based method. Using these MDs, we constructed predictive models using DataRobot. The calculated AUC and BAC were, respectively, 0.931 and 0.805 for LD50, 0.919 and 0.849 for BBBP, and 0.900 and 0.807 for the CL pathway. In this investigation, we constructed predictive models combining the DeepSnap-DL and MD-based methods. In ensemble models using the mean predictive probability of the DeepSnap-DL and MD-based methods, the calculated AUC and BAC were, respectively, 0.942 and 0.842 for LD50, 0.936 and 0.853 for BBBP, and 0.908 and 0.832 for the CL pathway, with improved predictive performance observed for all variables compared with either single method alone. Moreover, in consensus models that adopted only compounds for which the results of the two methods agreed, the calculated BAC for LD50, BBBP, and the CL pathway were 0.916, 0.918, and 0.847, respectively, indicating higher predictive performance than the ensemble models for all three variables. The predictive models combining the DeepSnap-DL and MD-based methods displayed high predictive performance for LD50, BBBP, and the CL pathway. Therefore, the application of this approach to prediction targets in various drug discovery screenings is expected to accelerate drug discovery.

Identification and Prioritization of Environmental Organic Pollutants: From an Analytical and Toxicological Perspective

Exposure to environmental organic pollutants has triggered significant ecological impacts and adverse health outcomes, which have been received substantial and increasing attention. The contribution of unidentified chemical components is considered as the most significant knowledge gap in understanding the combined effects of pollutant mixtures. To address this issue, remarkable analytical breakthroughs have recently been made. In this review, the basic principles on recognition of environmental organic pollutants are overviewed. Complementary analytical methodologies (i.e., quantitative structure-activity relationship prediction, mass spectrometric nontarget screening, and effect-directed analysis) and experimental platforms are briefly described. The stages of technique development and/or essential parts of the analytical workflow for each of the methodologies are then reviewed. Finally, plausible technique paths and applications of the future nontarget screening methods, interdisciplinary techniques for achieving toxicant identification, and burgeoning strategies on risk assessment of chemical cocktails are discussed.

International regulatory uses of acute systemic toxicity data and integration of new approach methodologies

Chemical regulatory authorities around the world require systemic toxicity data from acute exposures via the oral, dermal, and inhalation routes for human health risk assessment. To identify opportunities for regulatory uses of non-animal replacements for these tests, we reviewed acute systemic toxicity testing requirements for jurisdictions that participate in the International Cooperation on Alternative Test Methods (ICATM): Brazil, Canada, China, the European Union, Japan, South Korea, Taiwan, and the USA. The chemical sectors included in our review of each jurisdiction were cosmetics, consumer products, industrial chemicals, pharmaceuticals, medical devices, and pesticides. We found acute systemic toxicity data were most often required for hazard assessment, classification, and labeling, and to a lesser extent quantitative risk assessment. Where animal methods were required, animal reduction methods were typically recommended. For many jurisdictions and chemical sectors, non-animal alternatives are not accepted, but several jurisdictions provide guidance to support the use of test waivers to reduce animal use for specific applications. An understanding of international regulatory requirements for acute systemic toxicity testing will inform ICATM's strategy for the development, acceptance, and implementation of non-animal alternatives to assess the health hazards and risks associated with acute toxicity.

Toxicity Tests for Chemical Pesticide Registration: Requirement Differences among the United States, the European Union, Japan, and China?

Pesticide registration is a scientific, legal, and administrative process that checks if a pesticide is safe and effective for its intended use before it can be used. The toxicity test is a key part of pesticide registration, which includes human health and ecological effect testing. Different countries adopt their own toxicity test criteria for pesticide registration guidelines. However, these differences, which may help accelerate the progress of pesticide registration and reduce the number of animals used, are yet to be explored and compared. Herein, we outlined the details and compared the differences between the toxicity tests in the United States, the European Union, Japan, and China. Some differences lie in the types and waiver policy, while others are in new approach methodologies (NAMs). On the basis of these differences, there is great potential for the optimization of NAMs during the toxicity tests. It is expected that this perspective can contribute to developing and adopting NAMs.

Intelligent Workflow and Software for Non-Target Analysis of Complex Samples Using a Mixture of Toxic Transformation Products of Unsymmetrical Dimethylhydrazine as an Example

Unsymmetrical dimethylhydrazine (UDMH) is a widely used rocket propellant. Entering the environment or being stored in uncontrolled conditions, UDMH easily forms an enormous variety (at least many dozens) of transformation products. Environmental pollution by UDMH and its transformation products is a major problem in many countries and across the Arctic region. Unfortunately, previous works often use only electron ionization mass spectrometry with a library search, or they consider only the molecular formula to propose the structures of new products. This is quite an unreliable approach. It was demonstrated that a newly proposed artificial intelligence-based workflow allows for the proposal of structures of UDMH transformation products with a greater degree of certainty. The presented free and open-source software with a convenient graphical user interface facilitates the non-target analysis of industrial samples. It has bundled machine learning models for the prediction of retention indices and mass spectra. A critical analysis of whether a combination of several methods of chromatography and mass spectrometry allows us to elucidate the structure of an unknown UDMH transformation product was provided. It was demonstrated that the use of gas chromatographic retention indices for two stationary phases (polar and non-polar) allows for the rejection of false candidates in many cases when only one retention index is not enough. The structures of five previously unknown UDMH transformation products were proposed, and four previously proposed structures were refined.

Predicted health and environmental hazards of liquid crystal materials via quantitative structure-property relationship modeling

Liquid crystal materials (LCMs) are considered as emerging contaminants with high persistent and bioaccumulative potentials, but their toxicological effects are not well understood. To address this issue, a list of 1431 LCMs commercially available in the market was established through literature reviews and surveys of LCM suppliers. Toxicological properties of 221 target LCMs were derived from the Classification and Labeling Inventory by the European Chemicals Agency. More than 80 % of target LCMs likely pose adverse effects on human health or aquatic ecosystems. Two quantitative structure-property relationship (QSPR) models developed from the toxicological properties of LCMs achieved approximately 90 % accuracy in external data sets. The probability-based approach was more efficient in defining the applicability domain for the QSPR models than a range- or distance-based approach. The highest accuracy was achieved for chemicals within the probability-based applicability domain. The QSPR models were applied to predict health and environmental hazards of 1210 LCMs that had not been notified to the Classification and Labeling Inventory, and 301 and 94 LCMs were recognized as posing potential hazards to human health and the environment, respectively. The present study highlights the potential detrimental effects of LCMs and offers a specific in silico technique for screening hazardous LCMs.

Quantifying Functional-Group-like Structural Fragments in Molecules and Its Applications in Drug Design

A functional group in a molecule is a structural fragment consisting of a few atoms or a single atom that imparts reactivity to a molecule. Hence, defining functional groups is crucial in chemistry to predict the properties and reactivities of molecules. However, there is no established method in the literature for defining functional groups based on reactivity parameters. In this work, we addressed this issue by designing a set of predefined structural fragments along with reactivity parameters like electron conjugation and ring strain. This approach uses bond orders and atom connectivities to quantify the presence of these fragments within an organic molecule based on a given input molecular coordinate. To assess the effectiveness of this approach, we performed a case study to show the benefits of using these newly designed structural fragments instead of traditional fingerprint-based methods for grouping potential COX1/COX2 inhibitors by screening an approved drug library against aspirin molecule. The structural fragment-based model for ternary classification of rat oral LD₅₀ of chemicals showed performance similar to the fingerprint-based models. In evaluating the regression model performance for aqueous solubility, log(S), predictions, our approach outperformed the fingerprint-based model.

AI in drug discovery: A wake-up call

Following a proof-of-concept presentation on dual-use artificial intelligence (AI) in drug discovery by Collaborations Pharmaceuticals Inc. to the Swiss Federal Institute for NBC-Protection, we explored how a generative algorithm could develop the nerve agent VX and tens of thousands of analogs in a highly impactful Nature Machine Intelligence commentary. We not only laid out the experiment, but, with guidance from experts on arms control and dual-use, we called for more discussion around the general repurposing potential of AI in drug discovery. To continue that conversation, we now share further details on the experiment and place our experiences in the larger frame of other scientists who have similarly developed powerful technologies but without engaging with, or even truly understanding, the misuse potential and downstream consequences of the technologies. It is our sincere hope that our experiment may serve as an important wake-up call for users of generative AI.

Traditional Machine and Deep Learning for Predicting Toxicity Endpoints

Molecular structure property modeling is an increasingly important tool for predicting compounds with desired properties due to the expensive and resource-intensive nature and the problem of toxicity-related attrition in late phases during drug discovery and development. Lately, the interest for applying deep learning techniques has increased considerably. This investigation compares the traditional physico-chemical descriptor and machine learning-based approaches through autoencoder generated descriptors to two different descriptor-free, Simplified Molecular Input Line Entry System (SMILES) based, deep learning architectures of Bidirectional Encoder Representations from Transformers (BERT) type using the Mondrian aggregated conformal prediction method as overarching framework. The results show for the binary CATMoS non-toxic and very-toxic datasets that for the former, almost equally balanced, dataset all methods perform equally well while for the latter dataset, with an 11-fold difference between the two classes, the MolBERT model based on a large pre-trained network performs somewhat better compared to the rest with high efficiency for both classes (0.93-0.94) as well as high values for sensitivity, specificity and balanced accuracy (0.86-0.87). The descriptor-free, SMILES-based, deep learning BERT architectures seem capable of producing well-balanced predictive models with defined applicability domains. This work also demonstrates that the class imbalance problem is gracefully handled through the use of Mondrian conformal prediction without the use of over- and/or under-sampling, weighting of classes or cost-sensitive methods.

Advancing New Approach Methodologies (NAMs) for Tobacco Harm Reduction: Synopsis from the 2021 CORESTA SSPT-NAMs Symposium

New approach methodologies (NAMs) are emerging chemical safety assessment tools consisting of in vitro and in silico (computational) methodologies intended to reduce, refine, or replace (3R) various in vivo animal testing methods traditionally used for risk assessment. Significant progress has been made toward the adoption of NAMs for human health and environmental toxicity assessment. However, additional efforts are needed to expand their development and their use in regulatory decision making. A virtual symposium was held during the 2021 Cooperation Centre for Scientific Research Relative to Tobacco (CORESTA) Smoke Science and Product Technology (SSPT) conference (titled "Advancing New Alternative Methods for Tobacco Harm Reduction"), with the goals of introducing the concepts and potential application of NAMs in the evaluation of potentially reduced-risk (PRR) tobacco products. At the symposium, experts from regulatory agencies, research organizations, and NGOs shared insights on the status of available tools, strengths, limitations, and opportunities in the application of NAMs using case examples from safety assessments of chemicals and tobacco products. Following seven presentations providing background and application of NAMs, a discussion was held where the presenters and audience discussed the outlook for extending the NAMs toxicological applications for tobacco products. The symposium, endorsed by the CORESTA In Vitro Tox Subgroup, Biomarker Subgroup, and NextG Tox Task Force, illustrated common ground and interest in science-based engagement across the scientific community and stakeholders in support of tobacco regulatory science. Highlights of the symposium are summarized in this paper.

Bayesian inference of chemical exposures from NHANES urine biomonitoring data

BACKGROUND

Knowing which environmental chemicals contribute to metabolites observed in humans is necessary for meaningful estimates of exposure and risk from biomonitoring data.

OBJECTIVE

Employ a modeling approach that combines biomonitoring data with chemical metabolism information to produce chemical exposure intake rate estimates with well-quantified uncertainty.

METHODS

Bayesian methodology was used to infer ranges of exposure for parent chemicals of biomarkers measured in urine samples from the U.S population by the National Health and Nutrition Examination Survey (NHANES). Metabolites were probabilistically linked to parent chemicals using the NHANES reports and text mining of PubMed abstracts.

RESULTS

Chemical exposures were estimated for various population groups and translated to risk-based prioritization using toxicokinetic (TK) modeling and experimental data. Exposure estimates were investigated more closely for children aged 3 to 5 years, a population group that debuted with the 2015-2016 NHANES cohort.

SIGNIFICANCE

The methods described here have been compiled into an R package, bayesmarker, and made publicly available on GitHub. These inferred exposures, when coupled with predicted toxic doses via high throughput TK, can help aid in the identification of public health priority chemicals via risk-based bioactivity-to-exposure ratios.

Principles and Procedures for Assessment of Acute Toxicity Incorporating In Silico Methods

Acute toxicity in silico models are being used to support an increasing number of application areas including (1) product research and development, (2) product approval and registration as well as (3) the transport, storage and handling of chemicals. The adoption of such models is being hindered, in part, because of a lack of guidance describing how to perform and document an in silico analysis. To address this issue, a framework for an acute toxicity hazard assessment is proposed. This framework combines results from different sources including in silico methods and in vitro or in vivo experiments. In silico methods that can assist the prediction of in vivo outcomes (i.e., LD50) are analyzed concluding that predictions obtained using in silico approaches are now well-suited for reliably supporting assessment of LD50-based acute toxicity for the purpose of GHS classification. A general overview is provided of the endpoints from in vitro studies commonly evaluated for predicting acute toxicity (e.g., cytotoxicity/cytolethality as well as assays targeting specific mechanisms). The increased understanding of pathways and key triggering mechanisms underlying toxicity and the increased availability of in vitro data allow for a shift away from assessments solely based on endpoints such as LD50, to mechanism-based endpoints that can be accurately assessed in vitro or by using in silico prediction models. This paper also highlights the importance of an expert review of all available information using weight-of-evidence considerations and illustrates, using a series of diverse practical use cases, how in silico approaches support the assessment of acute toxicity.

Machine learning-assisted non-target analysis of a highly complex mixture of possible toxic unsymmetrical dimethylhydrazine transformation products with chromatography-mass spectrometry

Unsymmetrical dimethylhydrazine (UDMH) is a toxic and environmentally hostile compound that was massively introduced to the environment during previous decades due to its use in the space and rocket industry. The compound forms multiple transformation products, and many of them are as dangerous as UDMH or even more dangerous. The danger includes, but is not limited to, acute toxicity, chronic health hazards, carcinogenicity, and environmental damage. UDMH transformation products are poorly investigated. In this work, the mixture formed by long storage of the waste that contained UDMH was studied. Even a preliminary screening of such a mixture is a complex task. It consists of dozens of compounds, and most of them are missing in chemical and spectral databases. The complete preparative separation of such a mixture is very laborious. We applied several methods of gas chromatography-mass spectrometry and liquid chromatography-mass spectrometry, and several machine learning and chemoinformatics methods to make a preliminary but informative screening of the mixture. Machine learning allowed predicting retention indices and mass spectra of candidate structures. The combination of various ion sources and a comparison of the observed with the predicted spectra and retention was used to propose confident structures for 24 compounds. It was demonstrated that neither high-resolution mass spectrometry nor mass spectral library matching is enough to elucidate the structures of unknown UDMH transformation products. At the same time, the use of machine learning and a combination of methods significantly improves the identification power. Finally, machine learning was applied to estimate the acute toxicity of the discovered compounds. It was shown that many of them are comparable to or even more toxic than UDMH itself. Such an extremely wide and still underestimated variety of easily formed derivatives of UDMH can lead to a significant underestimation of the potential hazard of this compound.

Machine learning and artificial intelligence in toxicological sciences

Machine learning and artificial intelligence approaches have revolutionized multiple disciplines, including toxicology. This review summarizes representative recent applications of machine learning and artificial intelligence approaches in different areas of toxicology, including physiologically based pharmacokinetic (PBPK) modeling, quantitative structure-activity relationship modeling for toxicity prediction, adverse outcome pathway analysis, high-throughput screening, toxicogenomics, big data and toxicological databases. By leveraging machine learning and artificial intelligence approaches, now it is possible to develop PBPK models for hundreds of chemicals efficiently, to create in silico models to predict toxicity for a large number of chemicals with similar accuracies compared to in vivo animal experiments, and to analyze a large amount of different types of data (toxicogenomics, high-content image data, etc.) to generate new insights into toxicity mechanisms rapidly, which was impossible by manual approaches in the past. To continue advancing the field of toxicological sciences, several challenges should be considered: (1) not all machine learning models are equally useful for a particular type of toxicology data, and thus it is important to test different methods to determine the optimal approach; (2) current toxicity prediction is mainly on bioactivity classification (yes/no), so additional studies are needed to predict the intensity of effect or dose-response relationship; (3) as more data become available, it is crucial to perform rigorous data quality check and develop infrastructure to store, share, analyze, evaluate, and manage big data; and (4) it is important to convert machine learning models to user-friendly interfaces to facilitate their applications by both computational and bench scientists.

Artificial Intelligence-Based Toxicity Prediction of Environmental Chemicals: Future Directions for Chemical Management Applications

Recently, research on the development of artificial intelligence (AI)-based computational toxicology models that predict toxicity without the use of animal testing has emerged because of the rapid development of computer technology. Various computational toxicology techniques that predict toxicity based on the structure of chemical substances are gaining attention, including the quantitative structure-activity relationship. To understand the recent development of these models, we analyzed the databases, molecular descriptors, fingerprints, and algorithms considered in recent studies. Based on a selection of 96 papers published since 2014, we found that AI models have been developed to predict approximately 30 different toxicity end points using more than 20 toxicity databases. For model development, molecular access system and extended-connectivity fingerprints are the most commonly used molecular descriptors. The most used algorithm among the machine learning techniques is the random forest, while the most used algorithm among the deep learning techniques is a deep neural network. The use of AI technology in the development of toxicity prediction models is a new concept that will aid in achieving a scientific accord and meet regulatory applications. The comprehensive overview provided in this study will provide a useful guide for the further development and application of toxicity prediction models.

Predicting Prenatal Developmental Toxicity Based On the Combination of Chemical Structures and Biological Data

For hazard identification, classification, and labeling purposes, animal testing guidelines are required by law to evaluate the developmental toxicity potential of new and existing chemical products. However, guideline developmental toxicity studies are costly, time-consuming, and require many laboratory animals. Computational modeling has emerged as a promising, animal-sparing, and cost-effective method for evaluating the developmental toxicity potential of chemicals, such as endocrine disruptors, without the use of animals. We aimed to develop a predictive and explainable computational model for developmental toxicants. To this end, a comprehensive dataset of 1244 chemicals with developmental toxicity classifications was curated from public repositories and literature sources. Data from 2140 toxicological high-throughput screening assays were extracted from PubChem and the ToxCast program for this dataset and combined with information about 834 chemical fragments to group assays based on their chemical-mechanistic relationships. This effort revealed two assay clusters containing 83 and 76 assays, respectively, with high positive predictive rates for developmental toxicants identified with animal testing guidelines (PPV = 72.4 and 77.3% during cross-validation). These two assay clusters can be used as developmental toxicity models and were applied to predict new chemicals for external validation. This study provides a new strategy for constructing alternative chemical developmental toxicity evaluations that can be replicated for other toxicity modeling studies.

Predicting chemical hazard across taxa through machine learning

We applied machine learning methods to predict chemical hazards focusing on fish acute toxicity across taxa. We analyzed the relevance of taxonomy and experimental setup, showing that taking them into account can lead to considerable improvements in the classification performance. We quantified the gain obtained throught the introduction of taxonomic and experimental information, compared to classification based on chemical information alone. We used our approach with standard machine learning models (K-nearest neighbors, random forests and deep neural networks), as well as the recently proposed Read-Across Structure Activity Relationship (RASAR) models, which were very successful in predicting chemical hazards to mammals based on chemical similarity. We were able to obtain accuracies of over 93% on datasets where, due to noise in the data, the maximum achievable accuracy was expected to be below 96%. The best performances were obtained by random forests and RASAR models. We analyzed metrics to compare our results with animal test reproducibility, and despite most of our models "outperform animal test reproducibility" as measured through recently proposed metrics, we showed that the comparison between machine learning performance and animal test reproducibility should be addressed with particular care. While we focused on fish mortality, our approach, provided that the right data is available, is valid for any combination of chemicals, effects and taxa.

Predicting chemical ecotoxicity by learning latent space chemical representations

In silico prediction of chemical ecotoxicity (HC₅₀) represents an important complement to improve in vivo and in vitro toxicological assessment of manufactured chemicals. Recent application of machine learning models to predict chemical HC₅₀ yields variable prediction performance that depends on effectively learning chemical representations from high-dimension data. To improve HC₅₀ prediction performance, we developed an autoencoder model by learning latent space chemical embeddings. This novel approach achieved state-of-the-art prediction performance of HC₅₀ with R² of 0.668 ± 0.003 and mean absolute error (MAE) of 0.572 ± 0.001, and outperformed other dimension reduction methods including principal component analysis (PCA) (R² = 0.601 ± 0.031 and MAE = 0.629 ± 0.005), kernel PCA (R² = 0.631 ± 0.008 and MAE = 0.625 ± 0.006), and uniform manifold approximation and projection dimensionality reduction (R² = 0.400 ± 0.008 and MAE = 0.801 ± 0.002). A simple linear layer with chemical embeddings learned from the autoencoder model performed better than random forest (R² = 0.663 ± 0.007 and MAE = 0.591 ± 0.008), fully connected neural network (R² = 0.614 ± 0.016 and MAE = 0.610 ± 0.008), least absolute shrinkage and selection operator (R² = 0.617 ± 0.037 and MAE = 0.619 ± 0.007), and ridge regression (R² = 0.638 ± 0.007 and MAE = 0.613 ± 0.005) using unlearned raw input features. Our results highlighted the usefulness of learning latent chemical representations, and our autoencoder model provides an alternative approach for robust HC₅₀ prediction.

Evaluation of Variability across Rat Acute Oral Systemic Toxicity Studies

Regulatory agencies rely upon rodent in vivo acute oral toxicity data to determine hazard categorization, require appropriate precautionary labeling, and perform quantitative risk assessments. As the field of toxicology moves toward animal-free new approach methodologies (NAMs), there is a pressing need to develop a reliable, robust reference data set to characterize the reproducibility and inherent variability in the in vivo acute oral toxicity test method, which would serve to contextualize results and set expectations regarding NAM performance. Such a data set is also needed for training and evaluating computational models. To meet these needs, rat acute oral LD₅₀ data from multiple databases were compiled, curated, and analyzed to characterize variability and reproducibility of results across a set of up to 2441 chemicals with multiple independent study records. Conditional probability analyses reveal that replicate studies only result in the same hazard categorization on average at 60% likelihood. Although we did not have sufficient study metadata to evaluate the impact of specific protocol components (eg, strain, age, or sex of rat, feed used, treatment vehicle, etc.), studies were assumed to follow standard test guidelines. We investigated, but could not attribute, various chemical properties as the sources of variability (ie, chemical structure, physiochemical properties, functional use). Thus, we conclude that inherent biological or protocol variability likely underlies the variance in the results. Based on the observed variability, we were able to quantify a margin of uncertainty of ±0.24 log₁₀ (mg/kg) associated with discrete in vivo rat acute oral LD₅₀ values.

Machine learning models on chemical inhibitors of mitochondrial electron transport chain

Chemicals can induce adverse effects in humans by inhibiting mitochondrial electron transport chain (ETC) such as disrupting mitochondrial membrane potential, enhancing oxidative stress and causing some diseases. Thus, identifying ETC inhibitors (ETCi) is important to chemical risk assessment and protecting the public health. However, it is not feasible to identify all ETCi with experimental methods. Quantitative structure-activity relationship (QSAR) modeling is a promising method to rapidly and effectively identify ETCi. In this study, QSAR models for predicting ETCi were developed using machine learning methods. A clustering-based under-sampling (CBUS) method was developed to handle the imbalance issue in training sets. Structure-activity landscapes were generated and analyzed for training sets generated by the CBUS method. The consensus QSAR models constructed with CBUS achieved satisfactory performances (balanced accuracy = 0.852) in 100 iterations of five-fold cross validations, indicating the models can effectively classify ETCi. The classification model was further employed to screen chemicals in the Inventory of Existing Chemical Substances of China and 13 chemicals were identified as ETCi. Fifteen structural alerts for ETCi were identified in this study. These results demonstrated that the model and structural alerts are useful to screen ETCi.

Mapping Mechanistic Pathways of Acute Oral Systemic Toxicity Using Chemical Structure and Bioactivity Measurements

Regulatory agencies around the world have committed to reducing or eliminating animal testing for establishing chemical safety. Adverse outcome pathways can facilitate replacement by providing a mechanistic framework for identifying the appropriate non-animal methods and connecting them to apical adverse outcomes. This study separated 11,992 chemicals with curated rat oral acute toxicity information into clusters of structurally similar compounds. Each cluster was then assigned one or more ToxCast/Tox21 assays by looking for the minimum number of assays required to record at least one positive hit call below cytotoxicity for all acutely toxic chemicals in the cluster. When structural information is used to select assays for testing, none of the chemicals required more than four assays and 98% required two assays or less. Both the structure-based clusters and activity from the associated assays were significantly associated with the GHS toxicity classification of the chemicals, which suggests that a combination of bioactivity and structural information could be as reproducible as traditional in vivo studies. Predictivity is improved when the in vitro assay directly corresponds to the mechanism of toxicity, but many indirect assays showed promise as well. Given the lower cost of in vitro testing, a small assay battery including both general cytotoxicity assays and two or more orthogonal assays targeting the toxicological mechanism could be used to improve performance further. This approach illustrates the promise of combining existing in silico approaches, such as the Collaborative Acute Toxicity Modeling Suite (CATMoS), with structure-based bioactivity information as part of an efficient tiered testing strategy that can reduce or eliminate animal testing for acute oral toxicity.

Dual Use of Artificial Intelligence-powered Drug Discovery

An international security conference explored how artificial intelligence (AI) technologies for drug discovery could be misused for de novo design of biochemical weapons. A thought experiment evolved into a computational proof.

STopTox: An in Silico Alternative to Animal Testing for Acute Systemic and Topical Toxicity

BACKGROUND

Modern chemical toxicology is facing a growing need to Reduce, Refine, and Replace animal tests (Russell 1959) for hazard identification. The most common type of animal assays for acute toxicity assessment of chemicals used as pesticides, pharmaceuticals, or in cosmetic products is known as a "6-pack" battery of tests, including three topical (skin sensitization, skin irritation and corrosion, and eye irritation and corrosion) and three systemic (acute oral toxicity, acute inhalation toxicity, and acute dermal toxicity) end points.

METHODS

We compiled, curated, and integrated, to the best of our knowledge, the largest publicly available data sets and developed an ensemble of quantitative structure-activity relationship (QSAR) models for all six end points. All models were validated according to the Organisation for Economic Co-operation and Development (OECD) QSAR principles, using data on compounds not included in the training sets.

RESULTS

In addition to high internal accuracy assessed by cross-validation, all models demonstrated an external correct classification rate ranging from 70% to 77%. We established a publicly accessible Systemic and Topical chemical Toxicity (STopTox) web portal (https://stoptox.mml.unc.edu/) integrating all developed models for 6-pack assays.

CONCLUSIONS

We developed STopTox, a comprehensive collection of computational models that can be used as an alternative to in vivo 6-pack tests for predicting the toxicity hazard of small organic molecules. Models were established following the best practices for the development and validation of QSAR models. Scientists and regulators can use the STopTox portal to identify putative toxicants or nontoxicants in chemical libraries of interest. https://doi.org/10.1289/EHP9341.

Structures of Endocrine-Disrupting Chemicals Correlate with the Activation of 12 Classic Nuclear Receptors

Endocrine-disrupting chemicals (EDCs) can inadvertently interact with 12 classic nuclear receptors (NRs) that disrupt the endocrine system and cause adverse effects. There is no widely accepted understanding about what structural features make thousands of EDCs able to activate different NRs as well as how these structural features exert their functions and induce different outcomes at the cellular level. This paper applies the hierarchical characteristic fragment methodology and high-throughput screening molecular docking to comprehensively explore the structural and functional features of EDCs for the 12 NRs based on more than 7000 chemicals from curated datasets. EDCs share three levels of key fragments. The primary and secondary fragments are associated with the binding of EDCs to four groups of receptors: steroidal nuclear receptors (SNRs, including androgen, estrogen, glucocorticoid, mineralocorticoid, and progesterone), retinoic acid receptors, thyroid hormone receptors, and vitamin D receptors. The tertiary fragments determine the activity type by interacting with two key locations in the ligand-binding domains of NRs (N-H5-H3-C and N-H7-H11-C for SNRs and N-H5-H5'-H2'-H3-C and N-H6'-H11-C for non-SNRs). The resulting compiled structural fragments of EDCs together with elucidated compound NR binding modes provide a framework for understanding the interactions between EDCs and NRs, facilitating faster and more accurate screening of EDCs for multiple NRs in the future.

Revealing Adverse Outcome Pathways from Public High-Throughput Screening Data to Evaluate New Toxicants by a Knowledge-Based Deep Neural Network Approach

Traditional experimental testing to identify endocrine disruptors that enhance estrogenic signaling relies on expensive and labor-intensive experiments. We sought to design a knowledge-based deep neural network (k-DNN) approach to reveal and organize public high-throughput screening data for compounds with nuclear estrogen receptor α and β (ERα and ERβ) binding potentials. The target activity was rodent uterotrophic bioactivity driven by ERα/ERβ activations. After training, the resultant network successfully inferred critical relationships among ERα/ERβ target bioassays, shown as weights of 6521 edges between 1071 neurons. The resultant network uses an adverse outcome pathway (AOP) framework to mimic the signaling pathway initiated by ERα and identify compounds that mimic endogenous estrogens (i.e., estrogen mimetics). The k-DNN can predict estrogen mimetics by activating neurons representing several events in the ERα/ERβ signaling pathway. Therefore, this virtual pathway model, starting from a compound's chemistry initiating ERα activation and ending with rodent uterotrophic bioactivity, can efficiently and accurately prioritize new estrogen mimetics (AUC = 0.864-0.927). This k-DNN method is a potential universal computational toxicology strategy to utilize public high-throughput screening data to characterize hazards and prioritize potentially toxic compounds.