ICCVAM Biennial Report 2020-2021

There is a pressing need to rapidly and accurately assess the safety of environmental chemicals and reduce the number of animals used in regulatory testing while still protecting wildlife. NICEATM and the ICCVAM Acute Toxicity Workgroup organized a global collaborative project to develop predictive in silico models of acute oral systemic toxicity potential. Participants from 35 international groups submitted a total of 139 models built using a data set of 11,992 chemicals split into training (75%) and evaluation (25%) sets. These crowdsourced models were developed for five endpoints identified as relevant to regulatory decision frameworks: (1) LD50 value, (2) EPA hazard categories, (3) GHS hazard categories, (4) very toxic chemicals (LD50 <50 mg/kg), and (5) non-toxic chemicals (LD50 >2,000 mg/kg). Predictions within the applicability domains of the submitted models were evaluated and combined into consensus predictions based on a weight-of-evidence approach. The result, the Collaborative Acute Toxicity Modeling Suite (CATMoS; Mansouri et al. 2021), leverages the strengths and overcomes the limitations of individual modeling approaches. The consensus model predictions are fully reproducible and demonstrate equivalent performance to replicate in vivo data considering the inherent variability, offering a strong potential replacement for animal testing. Based on these results, CATMoS predictions for 178 chemicals are currently being evaluated in comparison to rat acute oral toxicity tests from publicly available ecological risk assessments registered from 1998 to 2020. Findings will inform the potential for using CATMoS estimates to potentially replace data from rat acute oral toxicity studies for assessments of ecological risk (i.e., what assumptions might be made, under what conditions, and which chemicals).

The CompTox Chemicals Dashboard is the primary web-based application that provides access to data and algorithms from the EPA Center for Computational Toxicology and Exposure. Since April 2017, a total of 10 releases have increased the data available from an initial release of 560k data points to a total of 906k data points with the December 2021 release. There are now over 310 individual chemical lists available that include substances such as pesticides, disinfectant byproducts, and PFAS. The latest release represents a major architecture and design update, with data management based on a datamart-datahub technology, a rich API, and fresh user interface design. With this architecture, new applications using the same underpinning data and API will be easier to build and maintain long-term. The update also supported significant improvements in performance. For example, batch searches and downloads for 5000 chemicals now generally take less than 5 seconds compared with over a minute in the previous version. Both Generalized Read-Across (GenRA) and Abstract Sifter are now separate modules, allowing their use for querying chemicals beyond those present in the underlying database. The next release of data, planned for 2022, will expand the number of substances to over 1.2 million chemicals and will include updates to many of the existing chemical lists.

OPERA is a free and open-source/open-data suite of QSAR models developed to support a range of research and regulatory purposes. In addition to physicochemical and environmental fate properties, OPERA offers a number of models predicting absorption, distribution, metabolism, and excretion (ADME) endpoints that are important to physiologically based pharmacokinetic modeling and IVIVE studies. The OPERA ADME-related endpoints were added during 2020 and 2021 and include models for octanol/water partition and distribution coefficients, acidic dissociation, fraction of chemical unbound to plasma protein, intrinsic hepatic clearance, and Caco2 permeability. All OPERA models were built using curated data sets split into training and test sets and molecular descriptors developed from standardized QSAR-ready chemical structures. Modeling adhered to the five principles for QSAR model development adopted by OECD. These principles support development of scientifically valid, high accuracy models with minimal complexity that support mechanistic interpretation, when possible.

Existing OPERA models are updated regularly when new experimental data are available. Recently, the models for octanol/water partition coefficient, fraction unbound, and intrinsic clearance were updated with the latest publicly available data sets to improve their predictivity and applicability domain coverage. For consistency and transparency, OPERA also provides a tool for standardizing chemical structures, an estimate of prediction accuracy, an assessment of applicability domain, and incorporation of experimental values when available. Technical and performance details are described in OECD-compliant QSAR Model Reporting Format reports. OPERA predictions are available through the EPA CompTox Chemicals Dashboard and the NTP’s Integrated Chemical Environment. The OPERA application can also be downloaded from the NIEHS GitHub repository as a command-line or graphical user interface for Windows and Linux operating systems.

The NTP's Integrated Chemical Environment (ICE) provides data and tools to help develop, assess, and interpret chemical safety tests. Updates to ICE during 2020 and 2021 launched several new tools and implemented substantive updates to existing tools.

• The Search tool now allows search results to be sent directly to other ICE tools, the EPA CompTox Chemicals Dashboard, and the NTP Chemical Effects in Biological Systems database. Visualization features implemented in March 2021 help users explore query results in more detail.
• Improvements to the IVIVE tool allowed users the option to upload their own in vitro data for IVIVE analysis and their own in vivo data for visualizing comparisons to IVIVE predictions, as opposed to being limited to using data available in ICE. An inhalation model was added to the IVIVE tool in October 2020. The tool now also allows assay selection based on mode of action and user specification of whether experimental or predicted data are used for ADME parameters. Results can also be filtered by mode of action or toxicity endpoint annotation.
• New visualization features in the Chemical Characterization tool help users better understand the relationships between members of a chemical set and lets users explore chemical use categories from the EPA’s Consumer Products database.
• Curve Surfer, launched in March 2021, allows users to view and interact with concentration–response curves from curated HTS data.
• The PBPK tool, launched in March 2021, uses models from the EPA’s httk package to generate predictions of tissue-specific chemical concentration profiles following a dosing event.
• Chemical Quest, launched in June 2021, uses Saagar molecular fingerprints (Sedykh et al. 2021) to identify chemicals in the ICE database having similar structures to a query chemical, which can be entered using chemical identifiers or structure drawings.

Improvements to all tools during 2020 and 2021 include organization of data from the EPA ToxCast and the U.S. government’s interagency Tox21 programs into query setup menus based on mechanistic targets and modes of action. Other improvements included mapping of ToxCast and Tox21 assays to controlled terms from the NCI Metathesaurus; new tooltips and information buttons to help users set up queries; and the acceptance of a variety of chemical identifier types as query inputs.

In collaboration with scientists at the University of North Carolina at Chapel Hill and the Universidade Federal de Goiás, NICEATM scientists contributed to the development of a model to predict skin sensitization potential of chemicals for two assays, the human patch test and murine local lymph node assay, and implemented this model in a web portal (Braga et al. 2017). Pred-Skin v. 3.0 (Borba et al. 2021) revised and expanded the freely accessible web tool to integrate multiple QSAR models developed with in vitro, animal in vivo, and human predictive patch test data into a consensus naïve Bayes model that predicts human effects.

Scientists from NCATS, NIEHS, and NCI and collaborators developed computational methods to predict chemical activity for 59 acute systemic toxicity endpoints across multiple species, including 36 endpoints for which computational models had not been previously developed. Data used to develop the models were collected and curated from the ChemIDplus database for acute systemic toxicity and represents the largest publicly available such data set, covering over 80,000 compounds. These data were used for developing multiple single- and multitask models utilizing random forest, deep neural networks, convolutional, and graph convolutional neural network approaches. The paper describing the project (Jain et al. 2021) also reports the consensus models based on different multitask approaches. The curated data set and the developed models have been made publicly available to support regulatory and research applications.

To improve utility and interpretability of molecular structure-based predictive models, NIEHS scientists and collaborators developed a novel set of extensible chemistry-aware substructures, Saagar. The Saagar features were systematically identified based upon open-source literature highlighting relationships among substructural moieties, physicochemical properties, toxicological properties of molecules, and ADME properties. This development approach makes Saagar features more interpretable than standard molecular descriptor libraries. The Saagar substructures were evaluated for their performance in chemical characterization and read-across applications by comparing results with four publicly available fingerprint sets for three benchmark chemical sets including about 145,000 compounds (Sedykh et al. 2021). In 18 of the 20 comparisons, Saagar features performed better than the other fingerprint sets. Saagar features efficiently characterize diverse chemical collections, thus making them a better choice for building interpretable predictive in silico QSAR models and read-across protocols.

To support rapid chemical toxicity assessment and mechanistic hypothesis generation, EPA and NIEHS scientists developed an intuitive webtool allowing a user to identify target organs in the human body where a substance is predicted to be more likely to produce effects (Borrel et al. 2020). This tool, Tox21BodyMap, incorporates results of 9,270 chemicals tested in the United States federal Tox21 research consortium using 971 HTS assays whose targets were mapped onto human organs using organ-specific gene expression data. Via Tox21BodyMap's interactive tools, users can visualize chemical target specificity by organ system and implement different filtering criteria by changing gene expression thresholds and activity concentration parameters. Dynamic network representations, data tables, and plots with comprehensive activity summaries across all Tox21 HTS assay targets provide an overall picture of chemical bioactivity.

Assays used in Tox21 and other HTS programs rely on luciferase and fluorescence-based readouts. These can be susceptible to signal interference by certain chemical structures resulting in false positive outcomes. However, the Tox21 portfolio includes assays specifically designed to measure interference in the form of luciferase inhibition and autofluorescence via multiple wavelengths and under various conditions. EPA, NIEHS, and NCATS scientists applied multiple machine learning algorithms to predict assay interference based on molecular descriptors and chemical properties (Borrel et al. 2020a). The best performing predictive models were incorporated into a web-based tool called InterPred (Borrel et al. 2020b) that allows users to predict the likelihood of assay interference for any new chemical structure. InterPred increases confidence in HTS data by decreasing false positive testing results.

In May 2021, NICEATM entered into an agreement with consumer products company Unilever to collaboratively test and further develop their Skin Allergy Risk Assessment (SARA) predictive model (Reynolds et al. 2019). SARA is a computational model that uses a variety of input data to estimate a probability that a chemical will cause an allergic skin reaction in humans. NICEATM will test the SARA model using a variety of chemical data sets, including chemicals of interest to U.S. and international regulatory agencies. NICEATM and Unilever will also work together to expand the SARA model to include data generated by the NIEHS Division of the NTP. The intent is to make the SARA model openly available for public use along with other NICEATM predictive models. Availability of the SARA model will help further reduce animal use for the endpoint of skin sensitization and will improve upon existing efforts by providing points of departure for quantitative human risk assessment.

NIEHS scientists developed of a set of computational models to predict eye irritation and corrosion (Sedykh et al. 2022). The models were developed using a curated database of in vivo eye irritation studies from the scientific literature and stakeholder-provided data. The database contained over 500 unique substances, including many mixtures, tested at different concentrations. Substances were categorized according to GHS and EPA hazard classifications. Two modeling approaches were used to predict classification of mixtures. A conventional approach generated predictions based on the chemical structure of the most prominent component of the mixture. A mixture-based approach generated predictions by using weighted feature averaging to consider all known components in the mixture. Results suggest that these models are useful for screening compounds for eye irritation potential. Future efforts to increase the models’ utility will focus on expanding their applicability domains and using them in conjunction with other input variables (e.g., in vitro data) to establish defined approaches for eye irritation testing. 

To identify the potential for a chemical to be of concern to sensitive populations, it is important for NAMs to characterize variations in metabolism that can affect toxicity within a population of interest. NIEHS scientists are incorporating the effects of genetic variability on ADME into pharmacokinetic models. These models can be used to predict a tissue concentration from an external exposure (forward dosimetry) or estimate an external exposure that would result in a plasma or tissue concentration equivalent to an effective concentration in an in vitro assay (reverse dosimetry). The models use inputs from ADMET Predictor software (Simulations Plus, Inc.) to characterize what enzymes might be involved in a chemical’s metabolism and the structure and proportions of metabolites produced. Work is ongoing to identify polymorphisms in the genes coding for enzymes that might affect metabolism, characterize the polymorphisms’ prevalence within populations, and incorporate this information into IVIVE and PBPK models. The OPERA QSAR modeling suite will then be used to predict a range of resulting toxicities for chemical metabolites.

In June 2021, OECD issued Guideline 497, Defined Approaches on Skin Sensitisation, the first internationally harmonized guideline to describe a non-animal approach to predict skin sensitization potential. NICEATM scientists are developing an interactive web application that computationally applies the defined approaches outlined in Guideline 497 through a user-friendly interface. The user uploads data from the test methods used in the defined approach to the web application, which then generates skin sensitization predictions for the user’s chemicals of interest. The user selects the analysis variables, and the application dynamically provides feedback about the user’s data set to identify problematic data values. The application is still in development and will be available on the NTP website in 2022.

Intestinal absorption plays a role in the toxicity of chemicals. The human Caco-2 cell line, derived from human colon epithelial cancer cells, is currently used in an in vitro model of human intestinal absorption. However, there are several limitations to this method, including long cell culture times and inability to conduct HTS. NIEHS scientists are developing computational models for chemical absorption in human intestines that would allow for screening a wide range of chemicals for potential intestinal permeability, which would assist in evaluating the toxicity potential of these chemicals. To date, a data set containing over 4,500 unique chemicals has been compiled and prepared; development of a model is planned for 2022.

The potential harms from inhaling flavoring chemicals used in e‑cigarettes and byproducts of those chemicals have not been extensively studied. One mechanism of interest is DNA adduct formation, which may lead to carcinogenesis. Scientists at the FDA Center for Tobacco Products (Kang and Valerio 2020) determined that flavoring chemicals and their byproducts include alkenylbenzenes and aldehydes, both of which are known to form DNA adducts. Using in silico toxicology approaches, they conducted a structural similarity analysis and generated in silico model predictions of these chemicals for genotoxicity, mutagenicity, carcinogenicity, and skin sensitization. Good concordance was observed between DNA adduct formation and models predicting mammalian mutagenicity and skin sensitization for both chemical classes. On the other hand, different prediction profiles were observed for the two chemical classes for the endpoints of unscheduled DNA synthesis and bacterial mutagenicity. These results are likely due to the different mode of action between the two chemical classes, as aldehydes are direct acting agents, while alkenylbenzenes require bioactivation to form the electrophilic intermediates that in turn form DNA adducts. This study suggests that in silico prediction for the mouse lymphoma assay may serve as a surrogate endpoint to help predict DNA adduct formation for chemicals found in tobacco products such as flavors and byproducts.

Scientists at the FDA Center for Tobacco Products (Goel and Valerio 2020) utilized in silico tools to predict the mutagenic potential of chemicals in tobacco products and tobacco smoke in a validation test and in a separate screening test. Publicly available QSAR models were validated against 900 chemicals relevant to tobacco products for which experimental Ames mutagenicity data were available from public sources. The predictive performance of the individual and combined QSAR models was evaluated using various performance metrics. All the models performed well, predicting mutagens and nonmutagens with 75%-95% accuracy, 66%-94% sensitivity, and 73%-97% specificity. Subsequently, in a screening test, a combination of complementary structure–activity relationship and QSAR models was used to predict the mutagenicity of 6,820 chemicals cataloged in tobacco products and/or tobacco smoke. More than 1,200 of these chemicals were predicted to have mutagenic potential, with 900 potential mutagens in tobacco smoke. This research demonstrates the validity of in silico QSAR tools to predict mutagenicity of chemicals in tobacco products or tobacco smoke, and suggests QSAR models may be useful as screening tools for hazard identification to inform tobacco regulatory science.

Serious adverse health effects have been reported with the use of vaping products, including neurologic disorders and e-cigarette or vaping product use-associated lung injury. Vitamin E acetate, often added as a diluent to cannabis-containing products, has been linked to lung injury. FDA scientists explored the relationship of such additives to neurotoxicity. Literature searches were performed on vitamin E and vitamin E acetate-associated neurotoxicity. The literature review showed that the neurotoxic potential of inhalation exposures to these compounds in humans is unknown. The blood-brain barrier penetration potential of vitamin E and vitamin E acetate were evaluated using cheminformatic techniques. Physicochemical properties suggest that these compounds are lipophilic, and molecular weights indicate vitamin E and vitamin E acetate have the potential for blood-brain barrier permeability. Computational models also predict both compounds may cross the blood-brain barrier via passive diffusion. The literature search found no experimental nonclinical studies or clinical information on the neurotoxic potential of vitamin E via inhalation. However, neurotoxic effects of phenyl acetate, a pyrolysis by-product structurally analogous to vitamin E acetate, suggest that vitamin E acetate has potential for central nervous system impairment. Cheminformatic model predictions provide a theoretical basis for potential central nervous system permeability of these inhaled dietary ingredients, suggesting that they should be prioritized to evaluate for potential central nervous system hazard.

Addiction to nicotine found in tobacco products causes difficulty in quitting among users. Nicotinic acetylcholine receptors (nAChRs) are the physiological targets of nicotine and facilitate addiction to tobacco products, primarily the nAChR-α7 subtype. Therefore, predicting the binding activity of tobacco constituents to nAChR-α7 is an important component for assessing addictive potential of tobacco constituents. Scientists at the FDA Center for Tobacco Products developed a predictive computational model for nAChR-α7 binding activity. The model was trained on data from 843 chemicals with human nAChR-α7 binding activity extracted from PubChem and the European Union’s ChEMBL chemicals database (Sakkiah et al. 2020). The model was tested using 1,215 chemicals with rat nAChR-α7 binding activity data from the same databases. The developed model was then used to predict the potential human nAChR-α7 binding activity for 5,275 tobacco constituents. The human binding activity data for 84 of the 5,275 tobacco constituents were experimentally measured to validate the prediction results. This model of human nAChR-α7 binding may be a useful tool for screening of potentially addictive tobacco constituents.

HTS approaches often employ microscopy to capture photomicrographs from multi-well cell culture plates, generating thousands of images that require manual human analysis. To automate this subjective and time-consuming process, NIEHS scientists and collaborators developed a method that uses deep learning to automatically classify digital assay images (Tandon et al. 2021). A convolutional neural network was trained to perform binary and multiclass classification. The binary classifier accurately binned assay images into healthy (comparable to untreated controls) and altered (not comparable to untreated controls) classes, while the multiclass classifier accurately assigned "Healthy," "Intermediate," and "Altered" labels to assay images. The study results indicated a strong correspondence between dosage and classifier-predicted scores, suggesting that these scores might be useful in further characterizing benchmark dose. Together, these results clearly demonstrate that deep learning-based automated image classification of cell morphology changes upon chemical-induced stress can yield highly accurate and reproducible assessments of cytotoxicity across a variety of cell types.

Traditionally, skin sensitization potential of chemicals has been determined using in vivo methods. One of these methods, the mouse lymph node assay, expresses the potential of a chemical to cause skin sensitization as an EC3 value: the concentration of a chemical required to elicit a threshold positive response. Scientists at the U.S. Air Force Research Laboratory explored approaches that use in silico and in vitro assays (Naumann and Arnold 2019) to calculate EC3 values and develop safe surface levels for chemicals used in fabrication and maintenance shops. A list of 125 relevant chemicals was compiled for evaluation of their skin sensitization potential using this approach. Of those chemicals, 21 known sensitizers have been identified, and an additional 58 chemicals were identified as having insufficient information available to determine sensitization potential. To support the analysis, additional in vitro testing was completed on a number of chemicals using two assays for skin sensitization potential, the direct peptide reactivity assay and the KeratinoSens assay. QSAR analysis using a nearest-neighbors approach was performed on the known sensitizers and measured or predicted EC3 values were obtained using Derek Nexus software (Lhasa Limited). Of the 58 chemicals lacking sensitization or toxicity data, predictions of sensitizing potential were also obtained using both Derek Nexus and publicly available OECD QSAR Toolbox software and results compared with six chemicals identified as potential sensitizers. Very few differences in sensitizer vs. nonsensitizer predictions of sensitization potential were found. The EC3 values derived from these efforts will be used to calculate surface levels. In vitro test results generated for the project will also be made available to update current in silico models of skin sensitization.

To fully characterize the potential human health risk of a substance, data are often needed on that substance’s toxicokinetics: how a substance is absorbed, distributed, metabolized, and eliminated in the body. Traditional approaches for obtaining toxicokinetics data use animals, but alternative approaches are being developed to computationally estimate toxicokinetic parameters. Scientists within EPA and the U.S. Air Force Research Laboratory are developing such an approach to generate high-throughput toxicokinetic estimates of chemical inhalation exposures (Linakis et al. 2020). This approach has been implemented in an inhalation component for the R-based high-throughput toxicokinetics (httk) package. In total, 142 exposure scenarios across 41 volatile organic chemicals were modeled and compared to published data. Ongoing work is focused on development of inhalation models for aerosols and mixed exposures. To better model and evaluate fetal exposure to environmental chemicals, EPA scientists have also developed a gestational model component for httk, which estimates the chemical concentrations in multiple fetal tissues (Kapraun et al. 2019, Kapraun et al. 2022). The gestational model has been used for IVIVE analysis by NICEATM to evaluate fetal exposures that cause developmental toxicity risk. NICEATM scientists have also actively participated in testing of httk and review of its code, data, and documentation.

To fully characterize the potential human health risk of a substance, data are often needed about a substance’s toxicokinetics to understand how it is absorbed, distributed, metabolized, and eliminated in the body. Traditional approaches for obtaining toxicokinetics data use animals, but alternative approaches are being developed to computationally estimate toxicokinetic parameters in silico using QSAR models. Scientists at the U.S. Air Force Research Laboratory (Sweeney and Sterner 2022) evaluated QSAR models that use chemical structure or property information to predict two toxicokinetic parameters: the maximal capacity for metabolism (Vmax) and the half-maximal concentration for metabolism (KM). None of the evaluated QSAR models in their published forms could be fully validated. Literature review, use of graphical information, and other strategies allowed the deficiencies to be addressed for QSAR predictions of these parameters for alkylbenzenes, volatile organic chemicals, and substrates of alcohol dehydrogenase, aldehyde dehydrogenase, cytochrome P450, and flavin-containing monooxygenases. The updated QSAR expressions for smaller data sets tended to show greater accuracy in Vmax and KM prediction than larger data sets, and Vmax was generally more accurately predicted than KM. In a feasibility case study, the QSAR models were used to estimate toxicokinetic parameters of jet fuel components to determine the potential utility of this approach for investigation of mixture toxicokinetics. Results suggested that the models performed better to predict toxicokinetic parameters of volatile organic chemicals and alkylbenzenes as compared to cytochrome P450 substrates, likely due to the physicochemical properties of the chemicals used in the models’ development.

Airmen are exposed to hundreds of chemicals in the operational environment. Some of these chemicals can interact with proteins in the brain and cause neurotoxic effects impairing memory and cognitive performance. To address the need for rapid assessment of neurotoxicity of chemicals, scientists at the U.S. Air Force Research Laboratory developed an in silico tool that uses a reverse molecular docking approach to identify receptor targets for chemicals (McCarthy et al. 2022). They also developed deep learning computational models that predict activity at the neurological targets. The chemical assessment process was automated via a Python-based graphical user interface that prepares all necessary files and performs docking and deep learning modeling. This is a novel tool that can potentially be used to screen and assess chemicals of interest for possible neurotoxic effects that can impact cognitive performance of airmen. 

Data gap filling techniques, such as QSAR models based on chemical structure information, can predict potential hazards of chemicals that have little experimental data. Risk assessment requires identification of a quantitative POD value, the lowest dose or concentration at which a treatment-related response is observed; effects of treatment by lower doses must be estimated by low-dose extrapolation. A study by EPA scientists (Pradeep et al. 2020) describes two sets of QSAR models to predict POD values for repeat-dose toxicity. The first set of QSAR models predicts point estimates of POD values using structural and physicochemical descriptors for repeat-dose study types and species combinations. The second set of QSAR models predicts the 95% confidence intervals for PODs using a constructed POD distribution based on previously published typical study-to-study variability that may lead to uncertainty in model predictions (Pham et al. 2020). Enrichment analysis to evaluate the accuracy of the predicted PODs showed that 80% of the 5% most potent chemicals were found in the top 20% of the most potent chemical predictions. This suggests that these repeat-dose POD QSAR models may help inform screening-level human health risk assessments in the absence of other data.