ICCVAM Biennial Report 2024-2025

Environmental incidents, such as jet fuel spills reported in different regions worldwide, can expose communities to harmful chemical mixtures through contaminated drinking water. However, testing the toxicity of such complex mixtures with traditional methods can be challenging when data on component chemicals are unavailable. ATSDR researchers explored the use of in silico tools, which rely on computer models and predictions, to estimate health risks from oral exposure to jet fuel (JP-5) mixtures (Ji et al. 2026). Publicly available in silico tools, such as threshold of toxicological concern estimates, quantitative structure–activity relationship (QSAR) models, and read-across approaches, were used to evaluate the potential health hazards of JP-5 and its chemical components. Applicability domain assessments indicated broad QSAR model coverage, with OPERA and ToxTree fully inclusive and VEGA models showing mostly moderate-to-high confidence for most additives. The models were used to estimate points-of-departure, which were then extrapolated using dosage adjustments and uncertainty factors to obtain health guidance values. The study found that in silico-derived health guidance values for JP-5 mixtures generally aligned with existing safety guideline values and that threshold of toxicological concern estimates were consistently conservative, which means they erred on the side of caution. Additionally, the researchers gathered data from previous studies to estimate possible health effects. Overall, the research demonstrated that advanced computer models can effectively fill in toxicological data gaps and provide valuable screening-level insights into the safety of chemical mixtures when traditional data is unavailable. This approach is a promising tool for addressing environmental health concerns and protecting public safety in situations involving chemical exposures.

Animal test data from evaluations of the toxicity of inhaled chemicals have limited applicability to humans due to anatomical differences between human and rodent respiratory tracts and complexities in the transport kinetics of inhaled chemicals through these organs. New approach methodologies (NAMs) have the potential to provide more human-relevant, mechanistically driven approaches to identify potential inhalation toxicity hazards. While in vitro assay data can provide valuable information about the effects of chemicals on biological systems, it is necessary to translate in vitro results to an in vivo context to determine safe exposure levels in real-life situations. In vitro to in vivo extrapolation (IVIVE) estimates the equivalent administered dose that would result in plasma concentrations equivalent to in vitro activity concentrations.

ATSDR and NICEATM scientists applied IVIVE to in vitro assay data from 20 volatile organic compounds with available pharmacokinetic and in vivo toxicity data covering multiple target organs via inhalation exposure. For each chemical, equivalent administered doses that would result in plasma or lung concentrations equivalent to the in vitro activity concentrations were estimated and compared to the in vivo points-of-departure (PODs) or ATSDR Minimal Risk Levels. While equivalent administered doses for most tested chemicals either underpredicted or were within 5-fold of in vivo PODs, substantial deviations from in vivo PODs warrant further investigation into the sources of uncertainty. However, most equivalent administered doses based on in vitro data were at least an order of magnitude higher than the corresponding Minimal Risk Levels, suggesting that an empirical or statistical modifying factor may be required to align equivalent administered dose-derived thresholds with public health-based guidance values. Physicochemical properties, particularly volatility, were found to significantly influence the reliability of in vitro assay results, and recognizing this and applying proper controls could improve results. There is also the possibility that incorporating experimental toxicokinetic data where available could improve the precision of IVIVE predictions and strengthen alignment with in vivo PODs. A paper describing this study is in preparation for submission in 2026.

EPA currently relies on animal testing, specifically rat studies, to determine the LD50 (the dose of a substance that causes death in 50% of test animals) for pesticide safety assessments. This information is used for pesticide toxicity labeling and environmental risk assessment. To reduce animal use, a consortium of international researchers led by NICEATM developed a computer-based tool called the Collaborative Acute Toxicity Modeling Suite (CATMoS; Mansouri et al. 2021). CATMoS uses QSAR modeling to predict acute oral toxicity (LD50) values for chemicals.

EPA and NIEHS researchers compared CATMoS predictions to real-world animal data for 177 pesticides (Bishop et al. 2024). The tool was evaluated for its ability to classify chemicals into the EPA’s acute toxicity categories and accurately predict LD50 values. Results showed that CATMoS was highly reliable at categorizing substances into EPA toxicity Categories III (moderate toxicity, 500–5000 mg/kg) and IV (low toxicity, over 5000 mg/kg), with an impressive 88% agreement between CATMoS predictions and animal test data for chemicals with LD50 values over 500 mg/kg. Additionally, for chemicals with LD50 values of 2000 mg/kg or higher, CATMoS predictions closely matched outcomes from high-dose animal tests.

These findings demonstrate that CATMoS is a reliable tool for predicting pesticide toxicity, particularly for less-toxic substances. By using CATMoS, regulators and industry can reduce the need for animal testing, making pesticide safety evaluations more ethical while still ensuring accurate risk assessments. This work represents a significant step forward in modernizing toxicity testing methods.

A study by the FDA National Center for Toxicological Research (NCTR) demonstrated how combining two analytical frameworks can improve predictions of drug toxicity while reducing reliance on animal models (Wang et al. 2025). The study compared and combined adverse outcome pathways (AOPs), which map how chemical exposures lead to harmful health effects through chains of biological events, and Bayesian networks, which are statistical models that calculate probabilities of outcomes based on interconnected factors. Both approaches use similar diagram structures called directed acyclic graphs, but researchers had not previously verified whether they are mathematically compatible.

Using drug-induced liver toxicity as a test case, the researchers confirmed that AOP networks align mathematically with Bayesian network properties. More importantly, they showed that applying specific mathematical concepts from Bayesian networks, particularly "Markov blankets" (the minimal set of factors needed to predict an outcome), can dramatically simplify toxicity prediction models. This simplification makes the models more efficient and practical for real-world use in drug development. The results support replacing traditional animal models with laboratory-based molecular assays that are more directly relevant to human biology. By demonstrating that these two frameworks work together effectively, the research provides a roadmap for developing better, faster, and more human-relevant methods to identify potentially toxic drugs earlier in the development process, ultimately improving drug safety while reducing animal use in testing.

Researchers at FDA NCTR compared traditional chemical analysis methods with super intelligence (SI) language models in the prediction of whether certain drugs and chemicals are toxic. Scientists tested whether SI models such as GPT-3 could predict toxicity as well as or better than established methods that analyze molecular structure (Gray and Wu 2025). Both approaches were applied to three different toxicity databases containing information about thousands of chemicals. The traditional molecular descriptor methods, which mathematically describe chemical structures, generally performed best on complex multi-endpoint toxicity predictions, achieving accuracy scores of approximately 0.855. However, SI language models showed superior performance for specific toxicity types, particularly clinical toxicity and drug-induced liver injury (DILI), with GPT-3 reaching near-perfect accuracy (0.996) for one dataset. The SI models analyzed three types of text inputs: chemical structure codes (SMILES strings), chemical names, and simple descriptions. Importantly, the researchers noted that the good performance of the SI models may partly reflect the models having encountered these well-known chemicals during training rather than true predictive ability. The study demonstrates that while traditional molecular analysis remains the gold standard for predicting the toxicity of new compounds, SI language models show promise as complementary tools, especially for screening known chemicals where existing toxicity information is available in scientific literature. Future work will combine both approaches to improve accuracy and reliability in drug safety evaluation.

Traditional research on the effects of chemicals on embryos and fetuses often relies on animal models, but use of human-relevant NAMs has the potential to improve safety assessments. An example is the Stemina devTOX quickPredict (devTOXqP) assay, which uses pluripotent stem cells to measure a chemical’s potential to cause developmental toxicity. In a preliminary study using the known developmental toxicant valproic acid and its analogs (Chang et al. 2022), NIEHS, EPA, FDA, and Stemina demonstrated that the devTOXqP assay, when used with advanced pharmacokinetic models, could estimate safe exposure limits and rank chemicals based on toxicity risks.

To better understand the applicability domain and performance of the devTOXqP assay, NICEATM and Stemina evaluated the accuracy of the assay in predicting in vivo developmental toxicity in humans and rodents for a broad range of 198 chemicals: 117 developmental toxicants, 67 developmental nontoxicants, and 14 chemicals with unknown developmental toxicity in vivo. For chemicals with the highest level of in vivo evidence, the devTOXqP assay performed well in predicting in vivo outcomes. The study results suggest that the devTOXqP assay may be useful for the identification of substances likely to cause perturbations in early development, with a negative result in the test considered along with other evidence in decisions to prioritize chemicals for in vivo testing. As the devTOXqP assay informs on a specific biochemical aspect of fetal development, it may also be useful as part of research strategies to clarify adverse outcome pathways for developmental neurotoxicity (DNT). A paper describing the study is in preparation for submission in 2026.

DNT occurs when exposure to a chemical causes an adverse effect in the developing brain or nervous system. NICEATM researchers and collaborators developed an approach that could be used to increase the speed and capacity of testing to predict whether a chemical might cause DNT without extensive animal testing (Kreutz et al. 2026). They used physiologically based pharmacokinetic (PBPK) models to translate doses that could potentially impair neurodevelopment from laboratory cell tests into estimates of human exposure levels that would be needed to disrupt brain development.

The team tested 91 chemicals that showed activity in cell-based tests for DNT, using three different PBPK modeling platforms. These PBPK models allowed for prediction of the amount of chemical that reaches the fetus and brain during critical periods of neurodevelopment. They focused on time points spanning pregnancy through six months after birth, when the brain is particularly vulnerable. The models predicted administered equivalent doses, the amount of chemical exposure that would be needed to produce concerning effects based on cell test results.

When compared against available animal study data for 31 chemicals, the predictions aligned well, with 78% falling within three-fold of observed effect levels. Data generated by the models were generally more protective compared with traditional animal model data. This work provides a faster, more cost-effective screening method for thousands of untested chemicals in commerce, helping regulators prioritize which substances warrant further investigation for potential DNT.

Traditional safety assessments often rely on generalized assumptions and standard uncertainty factors that may not adequately protect vulnerable groups, such as children, elderly individuals, or those with underlying health conditions. A review by EPA and NICEATM scientists highlights major advances in how differences across populations can be accounted for through physiologically based toxicokinetic modeling, which allows scientists to predict how chemicals move through and are processed by the human body (Kreutz et al. 2024).

Physiologically based toxicokinetic models incorporate detailed biological information about physiology and can be tailored to include age or population-related differences, such as differences in metabolism, genetic variability, and disease-related alterations in organ function. By integrating laboratory data from human cells and tissues with computational simulations, these models can translate in vitro findings into realistic human exposure scenarios. “Virtual populations” can also be modeled to simulate how diverse groups respond to chemical exposures, enabling more nuanced and inclusive risk assessments.

Applications of physiologically based toxicokinetic models for variability include evaluating lifetime exposures, chemical mixtures, and sensitive life stages, such as pregnancy. Regulatory agencies such as the FDA and EPA are increasingly using these tools to refine safety limits and improve decision-making. Overall, the body of work captured in this review represents a significant step toward personalized and population-informed risk assessment, improving public health protection while reducing reliance on animal models.

An NIEHS study combined biological measurements with IVIVE models in the NIH Integrated Chemical Environment to examine the real-world health impacts of triazole fungicides on coffee farmworkers and rural residents in Brazil (Marciano et al. 2024). Researchers detected multiple fungicides in urine samples from rural participants but not in samples from the urban control population, confirming that agricultural workers face significantly higher exposure rates. Importantly, male farmworkers showed reduced levels of key hormones including testosterone and androstenedione, which were associated with metabolic changes linked to increased risks of cardiovascular disease and diabetes. A notable aspect of the study was its integration of high-throughput laboratory data and computational toxicology tools into models to predict how these chemicals interact with human biological systems. These models successfully anticipated the hormonal disruptions and metabolic effects observed in the population, providing strong validation of nonanimal testing approaches. The findings raise particular concern about epoxiconazole, which showed effects at exposure levels experienced in the field. From a policy and regulatory perspective, this work demonstrates that modern computational methods can effectively assess real-world risks, supporting more proactive and protective safety standards. It also highlights the need for improved monitoring and regulation of pesticide exposure in agricultural settings, particularly for vulnerable worker populations.

Hearing loss can occur from exposure to certain medications or environmental chemicals, raising concerns about the possibility of substances causing toxicity to the auditory system, or ototoxicity. ATSDR researchers and collaborators used advanced in silico models to screen and prioritize environmental chemicals for their potential to cause auditory system toxicity, focusing on identifying hazards rather than calculating actual health risks (Luan et al. 2026). The study analyzed 80 environmental chemicals, excluding four already known to cause hearing damage. Researchers used computational methods, including structural similarity assessments and machine learning modeling, to predict auditory system toxicity risks. A consensus model identified 18 of the remaining 76 chemicals as potential auditory system toxicants with high accuracy. These data provide valuable insights for prioritizing these chemicals for further study. To explore mechanisms of auditory system toxicity, the researchers compared a polychlorinated biphenyl compound with an ototoxic drug called mitotane. Both substances were predicted to disrupt thyroid-stimulating hormone receptor signaling, suggesting that thyroid-related pathways may play a role in hearing damage. Additional molecular targets included the aryl hydrocarbon receptor, transthyretin, and the pregnane X receptor, which are linked to metabolic and inflammatory processes. These findings suggest common underlying pathways that may connect chemical exposure to hearing loss. The research highlights the power of computational modeling to support virtual screening and prioritization of chemicals for ototoxicity assessments, offering a quicker and more efficient approach to identifying risks and protecting public health, particularly for vulnerable populations.

Some chemicals used in food packaging and production may interfere with important genetic processes that regulate development and metabolism. A key concern centers on nuclear receptors, which control many biological processes, and retinoid-X receptors (RXRs). Disrupting RXRs can lead to harmful effects on multiple nuclear receptor pathways, impacting hormone function and metabolism.

To explore this risk, NICEATM researchers and collaborators used machine learning and computer simulations (in silico methods) to analyze over 57,000 chemicals and predict which would interact with RXR (Mengeling et al. 2025). Butylphenols emerged as the largest high-probability group, with 58 compounds identified, some of which are commonly used in food packaging. Further analysis of these chemicals using laboratory data from the EPA’s Toxicity Forecaster (ToxCast) program suggested that modifications to their molecular structure (specifically at the C4 location) could enable RXR activation.

To validate these predictions, researchers tested six types of butylphenols in amphibians (Xenopus laevis) using a model that measures thyroid hormone signaling, a pathway controlled by RXR. Three of these chemicals, with larger substitutions at C4, increased thyroid hormone activity at extremely low concentrations. Molecular modeling confirmed that bulkier chemical structures improved the fit and stability within the RXR binding pocket.

This study demonstrates how machine learning combined with in vivo experiments can effectively identify chemicals that may disrupt critical genetic pathways. The findings highlight the significance of chemicals used in food packaging, such as butylphenols, and provide tools for assessing their safety, ensuring better protection of human health and the environment.

Class-based methods, such as clustering and classification, are key to NAM development and application, aiding the understanding of hazard and risk concerns associated with groups of chemicals without additional laboratory work. Advances in computational chemistry, data generation and availability, and machine learning algorithms represent important opportunities for continued improvement of these techniques to optimize their utility for specific regulatory and research purposes. However, due to the intricacy of these methods, deep understanding and careful selection are imperative to align methods with their intended applications. 

NIEHS organized a 2022 workshop to consider available clustering and classification approaches and how they might be appropriately applied in a regulatory context. A subsequent review (Mansouri et al. 2024) provides important guidance on how scientists and regulators can more accurately group chemicals based on their similarities using clustering and classification approaches to predict potential health risks without extensive use of data from animal models. NIEHS researchers distinguish between two key approaches: general endpoint-agnostic similarity, which compares overall structural features such as shape and functional groups, and endpoint-specific similarity, which focuses on linking structural characteristics to particular biological effects such as toxicity or receptor binding. A central finding is that using endpoint-agnostic similarity methods when endpoint-specific predictions are required can lead to incorrect conclusions, as visibly similar chemical structures may behave very differently biologically. To address this, the team developed an automated and user-friendly workflow tool that helps nonexperts apply the correct analytical approach. This framework is particularly valuable for “read-across” strategies, where data from well-studied chemicals are used to predict the safety of similar but untested substances. By improving the reliability of these predictions, the work supports regulatory decision-making at agencies such as EPA and the European Chemicals Agency. Overall, this research advances NAMs by making sophisticated computational toxicology tools more accessible, helping reduce reliance on animal models while improving the scientific basis of chemical safety assessments.