ICCVAM Biennial Report 2024-2025

The CompTox Chemicals Dashboard is the primary web-based application for access to data and algorithms from the EPA Center for Computational Toxicology and Exposure. The Dashboard is a widely used resource for chemistry, toxicity, and exposure information for over a million chemicals. There were eight updates to the Dashboard released in 2024 and 2025, increasing the number of chemicals to 1,246,399 with all related data. These related data include mass spectrometry-ready/quantitative structure–activity relationship (QSAR)-ready structures, linked substances, and property predictions generated by the Open (Quantitative) Structure–activity/property Relationship App (OPERA) and EPA Toxicity Estimation Software Suite tools. The EPA Toxicity Forecaster (ToxCast) invitrodb v. 4.2 dataset was released in September 2024, and an updated version of the Chemical and Products Database was released via the Dashboard in June 2025 (Handa et al. 2025). The next Dashboard update is planned for August 2026. 

A June 2025 virtual training on the Dashboard gave an overview of Dashboard functions along with its risk evaluation components, navigation, and a case study. Additional training resources for the Dashboard are available on the EPA NAMs Training webpage.

EPA’s Sequence Alignment to Predict Across Species Susceptibility (SeqAPASS) is a fast online screening tool that allows researchers to extrapolate toxicity information across species.

There were four updates to SeqAPASS in 2024 and 2025, the most recent being the release of SeqAPASS v. 9.0 in September 2025. Improvements to SeqAPASS in these releases included:

• Protein structure generation and alignment capabilities to support detailed analyses of protein structural biology
• Incorporation of data from the World Register of Marine Species database to allow filtering of marine species
• Addition of a Focus/Filter toggle to customize results displays to either highlight optional selections from a full list of species or filter results to only display species that match the optional selection
• Data and executable updates

Training on new SeqAPASS capabilities was provided at an EPA virtual training session in May 2025. SeqAPASS v. 9.1 is scheduled for release in early 2026 and will include additional data and executable updates; v. 10.0 will be released later in 2026.

Ecological risk assessments aim to protect diverse ecological communities from chemical stressors through the evaluation of available species sensitivity data. However, toxicity data are typically available for only some standard test species, with limited or no data available for most species and taxa. Web-based Interspecies Correlation Estimation (Web-ICE), a web application developed by EPA, uses available data from appropriate surrogate species to predict acute toxicity for other species of interest lacking toxicity data. It can help address data gaps in species sensitivity and reduce reliance on uncertainty factors in ecological risk assessment. 

Web-ICE v. 4.0 was released in May 2024. This release has over 4,700 models representing 478 aquatic animal species, over 100 models representing 69 algal species, and over 850 models for 156 terrestrial bird and mammal species. A new feature of Web-ICE v. 4.0 is the Bulk ICE module, which generates predictions for all available taxa for each user-entered surrogate, maximizing taxonomic diversity in the resulting model output. Slides from a December 2024 webinar describing features of Web-ICE v. 4.0 are available on the EPA website, and a video is available on YouTube.

The NIH Integrated Chemical Environment (NICE), developed by NICEATM, provides data and tools to help develop, assess, and interpret chemical safety tests. All tool inputs now provide a “filter-as-you-type” function to help users find Chemical Quick Lists of interest. NICE tool help videos were updated in the September 2024 NICE v. 4.1 release and the February 2025 NICE v. 4.1.1 release. The updates also implemented improvements to many specific NICE tools.

• Major updates to Search result data visualizations in the March 2024 v. 4.0.2 release added immersive visualizations and functional use pagination; this release also updated filter chain options for Search results.
• The March 2024 update of the In Vitro to In Vivo Extrapolation (IVIVE) tool added new options for data overlays to contextualize model results. The July 2025 NICE v. 4.2 release also enabled users to provide their own physicochemical and absorption, distribution, metabolism, and excretion (ADME) data to Physiologically Based Pharmacokinetics (PBPK) and IVIVE tool queries.
• The July 2025 release revised the principal component analysis plots in the Chemical Characterization tool.
• Multiple improvements were made to the Curve Surfer tool, which allows users to view and interact with concentration–response curves from curated high-throughput screening (HTS) data. The tool now allows users to overlay AC50 or ACC values and includes technological interference flags and points-of-departure in results. Curve Surfer data are also now available via the NICE REST API.

NICEATM is working toward replacing traditional animal models with computational models to predict inhalation toxicity (LC50 values). The LC50 value represents the concentration of a substance that causes harm or death in half of a test population after inhalation and is an important measure used in chemical hazard assessment. NICEATM is developing a model to predict LC50 values that addresses an action item from a 2016 workshop focused on finding humane and efficient alternatives to animal tests for inhalation toxicity (Clippinger et al. 2018).

To support the project, NICEATM curated a dataset of LC50 values for roughly 760 chemicals to serve as the foundation for the development of predictive models. Unlike simpler endpoints, inhalation toxicity presents unique challenges due to complex data and regulatory requirements, making this initiative particularly innovative.

In mid-2024, collaboration with an international consortium of 35 participating groups began, with the goal of creating tailored computational tools specifically designed to predict inhalation toxicity. These models, combined into a consensus model and implemented in the free and open-source suite OPERA, will focus on meeting regulatory standards by classifying chemicals into hazard categories for systemic toxicity assessments. The collaborative process involves refining the methodologies and building consensus among researchers and regulators to ensure the tools are both scientifically robust and broadly applicable. The consensus model has been shared with project contributors. In 2026, developers will implement the consensus model in OPERA, making it publicly available. A paper on the modeling project will be submitted to a peer-reviewed journal.

This project represents a major step forward in toxicology, offering not only more humane testing strategies that eliminate reliance on animal studies but also efficient approaches to evaluating chemical safety. These computational models will help align with modern regulatory frameworks while advancing the science behind inhalation toxicity assessments.

Reducing, refining, and replacing the use of animals (applying 3Rs principles) is a long-standing goal in chemical toxicology. To support this effort, NICEATM and collaborators developed STopTox (Systemic and Topical Chemical Toxicity), an in silico platform that uses advanced computer models to predict chemical toxicity without the need for animal studies (Borba et al. 2022). STopTox uses QSAR models, which analyze chemical structures to estimate toxicity. Designed as a user-friendly, accessible tool, STopTox provides a comprehensive alternative for identifying hazards.

During its initial development, STopTox was rigorously tested using large publicly available datasets, ensuring it met international standards, including principles for QSAR model development adopted by the Organisation for Economic Co-operation and Development (OECD). The models demonstrated high internal accuracy and strong external predictive capability. To further build confidence in STopTox as a substitute for animal testing, researchers conducted external validation tests using independent datasets. They assessed the ability of STopTox to predict toxicity across six major categories: acute oral, dermal, and inhalation toxicity, as well as skin sensitization, skin irritation/corrosion, and eye irritation/corrosion. The validation results were highly encouraging, as STopTox performed reliably across all endpoints (Scheufen Tieghi et al. 2025). This supports the platform's potential as a dependable tool for regulatory toxicology decision-making. By offering accurate predictions and reducing the reliance on animal studies, STopTox represents an important step forward in advancing humane and efficient approaches to hazard identification in chemical testing. It highlights the promise of computational methods in revolutionizing toxicological assessments and contributing to animal replacement efforts.

There is a need for reliable and easy-to-use navigation tools capable of browsing, mining, and utilizing large and diverse repositories of chemical and biological data. To address this need, NIEHS in collaboration with Sciome, LLC, developed OrbiTox. OrbiTox uniquely offers an interactive and immersive 3D environment for the visualization of millions of chemical data entries, including both experimental and predicted biological activities toward various gene targets tested in vitro and/or in vivo. By organizing those data entries in concentric “orbits” (e.g., chemicals, genes, pathways, species), OrbiTox facilitates translational discovery by inferring knowledge from connections across multiple data domains and sources (Ross et al. 2025). Through an intuitive graphical user interface, OrbiTox offers rapid and interactive navigation in the chemical space along with extensive gap-filling capabilities. Currently, OrbiTox employs over 150 predictive and cross-validated QSAR models, including individual models for Tox21 assays, bacterial mutagenicity, cardiotoxicity failure modes, and carcinogenicity classification and potency. These models use Saagar molecular fingerprints (Sedykh et al. 2021), which provide interpretable, chemistry-backed reasoning for each prediction with state-of-the-art atom contributions. As a result, users can hypothesize on the substructural origins and/or mechanistic steps highlighting why a given chemical exhibits a particular activity. Such identification of favorable versus unfavorable chemotypes for a given biological property can be of critical importance for prioritizing experimental testing. OrbiTox also provides rich, structurally relevant analogs for read-across assessments, thanks to Saagar-based structural similarity and its vast database of bioactivity and metabolic profiles. Finally, OrbiTox provides users with opportunities to directly compare animal versus nonanimal data, which is important for building confidence in and establishing the relevance of new approach methodologies, ultimately supporting the transition to human-based safety testing. The fall 2026 release will include an extended set of new QSAR models and advanced chemical filtering features to further refine chemical search results and predictions.

NICEATM and collaborators at the German Federal Institute for Risk Assessment developed an approach that incorporates data from human predictive patch tests (HPPTs) to consider skin sensitization potency in assigning chemicals to subcategories of United Nations Globally Harmonized System of Classification and Labelling of Chemicals (GHS) hazard classification (Herzler et al. 2024). This approach to potency classification is potentially more human-relevant than the current classification approach.

While the modified classification approach is useful, it can be challenging to understand and implement. To address this barrier, NICEATM developed the HPPT App, which provides an interactive platform for users to input HPPT data, visualize results, and understand classification outputs. The app is publicly available on the NIH website with no registration required. 

Following the approach described by Herzler et al., the HPPT App uses data on both the number of sensitized human test subjects and the dose required to sensitize a single human test subject. If the data do not support a definitive classification, the HPPT App provides ambiguous classifications that indicate that there is either no evidence for the chemical being a strong sensitizer or that the data provided are insufficient to support any classification. The HPPT App includes a weight-of-evidence method to consolidate multiple HPPT results into a single GHS classification for each chemical. It flags questionable input data, and uploaded query data are not stored after the user ends their session. The app also provides a user guide and background information.

Regulatory agencies require that chemical products be evaluated for their potential to cause skin sensitization. Traditionally, such evaluations involved animal models, but more recently, nonanimal methods have been developed. Defined approaches for skin sensitization (DASS) use standardized interpretation procedures to combine data from multiple nonanimal tests, such as human cell-based experiments, molecular studies, and computational models, to determine if a chemical has potential to cause allergic dermatitis.

While regulatory agencies accept several DASS methods, applying these approaches manually can be complex, slow, and prone to errors. To address this, NICEATM researchers created the DASS App (To et al. 2024), a free, open-source web application that simplifies the use of three regulatory testing strategies: Two-out-of-Three, Integrated Testing Strategy, and Key Event 3/1 Sequential Testing Strategy. This web tool, first launched in 2023, allows users to upload their own data, identify inconsistencies, and generate predictions without needing advanced computational skills or special software installation. DASS App provides a data template and the option to select the assays from which the data are taken. Reference data can be imported from NICE. Outputs are provided in a user-friendly, downloadable format as tables and interactive graphics. By making the application freely accessible online, the developers ensure the tool is available for wide adoption and future improvements. This innovation supports faster, more reliable hazard assessments while advancing the goal of replacing the use of animals with humane and efficient alternatives.

Updates in 2024 and 2025 made the app more user-friendly, including simplified handling of borderline calls, as well as adding new test method inputs to cover similar updates to OECD Test Guideline 497, which contains the Two-out-of-Three and Integrated Testing Strategy defined approaches. DASS App now allows the user to evaluate performance using their own reference data or data from NICE chemical reference lists.

Assessing the potential of chemicals to cause skin sensitization (allergic reaction) is essential for protecting public health. While nonanimal methods have been developed, single tests alone are not sufficient to fully determine a chemical’s sensitization risk or potency. Defined approaches (DAs) such as those described in OECD TG 497 (OECD 2025) combine multiple data sources for more reliable evaluations. However, existing DAs lack the ability to calculate a specific dose threshold, known as a point-of-departure (POD), for use in risk assessments.

The Skin Allergy Risk Assessment–Integrated Chemical Environment (SARA-ICE) DA for skin sensitization (Reinke et al. 2025) was developed by NICEATM and the consumer products company Unilever specifically to address this challenge. SARA-ICE employs statistical modeling to estimate a human-relevant metric called the ED01. The ED01 is the probability distribution for the dermal dosage at which there is a 1% chance of inducing sensitization, making it directly applicable to risk assessment. SARA-ICE integrates data from various in vitro tests as well as data from the local lymph node assay and HPPT, drawing upon historical data from 434 chemicals, and accounts for variability and uncertainty in input data. In addition to the ED01, SARA-ICE provides probabilities for GHS skin sensitization classifications. Specifically, the model calculates the probability that a chemical would be placed into a GHS subcategory: severe sensitizers (1A), moderate sensitizers (1B), or Not Classified (NC). In a case study focused on isothiazolinones, SARA-ICE demonstrated its ability to integrate diverse data inputs and calculate ED01 values for six isothiazolinone compounds that compared favorably with results based on historical datasets (Reinke et al. 2025).

In 2025, OECD revised TG 497 to include a more limited guideline-specific version of the SARA-ICE model, which provides a POD as the geometric mean of the ED01 and does not provide hazard or GHS classification predictions. It also assumes that any chemical evaluated is presumed to be a skin sensitizer due to model restrictions.

To broaden the accessibility of the SARA-ICE model, in 2025 NICEATM developed the SARA-ICE app. The app provides access to the two SARA-ICE models: the "OECD TG 497 DA”, which aligns with TG 497 criteria, and the “Extended Model”, which allows the user to make a hazard determination and predict GHS subcategorization and alternative “EDx” outputs (e.g. ED05, the dose at which 5% of a population would be sensitized). The Extended Model does not assume a chemical is a skin sensitizer and allows the user to adjust the GHS decision threshold depending on desired stringency. The OECD TG 497 DA is static because it is based on the guideline; however, the Extended Model could be modified in the future to add in more test methods or outputs. Current outputs include a graphic representation of the uncertainty distribution and a tabular report of the probability distributions for various classifications. Future development may add methods to the Extended Model, in silico reactivity structural alerts, extrapolation of population-level effects, a user input feature, and a downloadable local version.

Skin sensitization, or an allergic reaction due to chemical exposure, is a critical focus in chemical safety assessments. Traditional animal tests often fail to accurately predict human responses, and ethical concerns further limit their use. To address this, NICEATM and University of North Carolina researchers developed HuSSPred, a reliable and ethical in silico tool designed using exclusively human data to predict human skin sensitization based on chemical structures. HuSSPred builds on HPPT data and aims to improve regulatory assessments while aligning with the 3Rs principles. The research team compiled an extensive HPPT database and grouped chemicals by their characteristics. They then developed and fine-tuned QSAR models using advanced optimization techniques. These models were rigorously tested using five-fold cross-validation and further validated on external data from the DASS App tool.

HuSSPred demonstrated strong predictive performance, with a correct classification rate ranging from 55%–88%, sensitivity (ability to correctly identify chemicals that would induce skin sensitization) between 48% and 89%, and specificity (ability to correctly identify chemicals that would not induce skin sensitization) between 37% and 92%. The tool’s positive predictive value was particularly robust, ranging from 84%–97%, indicating strong reliability in identifying sensitizing chemicals. Validation results showed performance comparable to or better than existing tools, confirming its usefulness for regulatory assessments. HuSSPred is publicly accessible through a user-friendly web interface, providing an effective alternative to traditional methods. Its high accuracy, ethical approach, and open access make it a valuable resource for predicting skin sensitization, supporting better regulatory decisions and reduced reliance on animal use.

A publication describing HuSSPred (Scheufen Tieghi et al. 2024) won the best paper of the year award from the Society of Toxicology In Vitro and Alternatives Specialty Section and was named an Editor's Choice publication in the journal MDPI Toxics.

Analyzing large chemical datasets to identify patterns and categorize chemicals into meaningful groups is essential for advancing chemical safety and developing innovative testing methods. Chemical grouping involves organizing chemicals based on shared characteristics, such as physicochemical properties, biological activity, or intended uses. However, many existing tools for chemical grouping require advanced programming knowledge or expensive commercial software, making them inaccessible to many users.

To address these challenges, NICEATM developed MoVIZ, a user-friendly and automated pipeline of workflows using KNIME, a free and open-source data analytics platform (Moreira-Filho et al. 2024). MoVIZ tools include a chemical grouping workflow that guides users through complex tasks, such as calculating molecular descriptors, grouping chemicals based on structural and biological similarities, and applying machine learning techniques to identify patterns. It thus provides a comprehensive, no-code solution for chemical grouping and visualization. This workflow includes tools for calculating molecular properties, selecting important features, reducing data complexity, and applying machine learning methods (both supervised and unsupervised) to group chemicals effectively. Additionally, the workflow can generate interactive visualizations and natural language summaries to explain why specific groupings were made, making the results clear and interpretable. A case study using a dataset of over 2,200 chemicals tested for eye irritation successfully identified clusters of irritants and nonirritants and highlighted the molecular features driving these effects. To further demonstrate MoVIZ’s capabilities, developers applied the workflow to a dataset focused on eye irritation and corrosion, highlighting its effectiveness in categorizing chemicals and identifying key molecular properties for each group.

Designed for accessibility, MoVIZ workflows can be applied seamlessly on both the KNIME desktop application and the KNIME Server WebPortal, functioning as an interactive web application with guided steps. MoVIZ tools enable a broad range of researchers, regulators, and industry professionals to apply modern computational toxicology methods. This supports more efficient chemical prioritization, safer product design, and reduced reliance on animal models, aligning with global efforts to modernize safety assessment practices. Such innovative workflows empower users without coding expertise to perform advanced chemical analyses, democratizing access to powerful data analysis tools. MoVIZ represents a significant step forward in enabling efficient, transparent, and reproducible chemical grouping, supporting regulatory decision-making and chemical safety assessments.

Understanding the safety of medications during pregnancy is critical, but many drugs are not adequately studied for their effects on the developing fetus. To help address this gap, NICEATM and University of North Carolina researchers developed advanced QSAR models to predict the likelihood of developmental toxicity from chemicals, including trimester-specific risks. These models aim to provide safer medication options for pregnant women while reducing the need for animal testing and aligning with 3Rs principles. The team collected and curated developmental toxicity data from FDA and Teratogen Information System databases, validating the information with thorough literature reviews. Using machine learning algorithms, they developed QSAR models that predict developmental toxicity with high accuracy: 76% accuracy for overall pregnancy risk and 80%, 95%, and 95% accuracy for risks specific to the first, second, and third trimesters, respectively. Sensitivity ranged from 53%–90%, while specificity varied between 46% and 100%, based on rigorous validation methods. These models have been made accessible through a public web platform called DeTox (Developmental Toxicity; Scheufen Tieghi et al. 2025). This tool supports regulatory evaluations of chemicals and drugs, helping the pharmaceutical and cosmetic industries design safer products for pregnant populations. Additionally, the curated dataset of developmental toxicants is available for public use. Overall, DeTox represents a significant step forward in improving prenatal safety assessments, reducing reliance on animals, and advancing the development of safer treatments for expectant mothers and their babies.

Access to visualization tools to navigate chemical space has become increasingly important due to the increasing size and diversity of publicly accessible compendiums of HTS and other descriptor and effects data. Construction of such tools relies on complex projection techniques using molecular descriptors. However, application of these techniques requires advanced programming skills that are beyond the capabilities of many stakeholders.

Inspired by the Google Maps application, NICEATM developed ChemMaps.com to easily navigate chemical space. The first version of ChemMaps.com was limited to exploration of drugs and drug candidates. ChemMaps.com v. 2.0, released in 2022, added data on approximately one million environmental chemicals from the EPA Distributed Structure-Searchable Toxicity (DSSTox) inventory (Borrel et al. 2023). ChemMaps.com v. 2.0 incorporated mapping to HTS assay data from the U.S. federal Tox21 research collaboration, which includes results from approximately 2,000 assays tested on up to 10,000 chemicals. Users can visualize chemical activity by both assay and target directly on the map and compare chemical spaces occupied by active and inactive chemicals. The latest update of ChemMaps.com (v. 2.1), released in October 2024, refined the chemical preparation pipeline to address previously unsupported chemistry. Chemical predictions were also updated to utilize the latest version of OPERA (v. 2.9).

If not evaluated or used properly, some chemicals in food may have toxic properties that pose risks to human health. FDA helps to safeguard the food supply by evaluating the safe use of chemicals as food ingredients and food contact substances. The FDA also monitors the food supply for chemical contaminants and takes action to limit exposure when the level of a contaminant could cause a food to be unsafe. To do this, FDA uses scientific and regulatory tools to assess the safety of chemicals in food while continually updating its methods to incorporate new approaches.

In July 2025, FDA released its Expanded Decision Tree chemical toxicity and risk screening tool that advances the agency’s commitment to transparency and enhancement of the food chemical safety program. The tool provides a consistent, systematic, and science-based approach to support evaluation of the safety of chemicals in food based on their structure and estimated toxicity. The Expanded Decision Tree was evaluated through external peer review, marking a significant milestone in its development. The update also incorporated input from stakeholders such as EPA, the National Toxicology Program, the European Food Safety Authority, and experts from academia and industry. FDA encourages interested parties to engage with the agency on the Expanded Decision Tree. Planned opportunities for public engagement include release of an informational video and listening sessions with technical, scientific, and trade groups to learn how the tool may impact their work.

Researchers at the FDA National Center for Toxicological Research (NCTR) have developed advanced super intelligence (SI) tools to enhance adverse event detection from diverse data sources, such as social media posts and FDA drug labeling documents. These projects demonstrated high accuracy in extracting and classifying adverse event data, supporting faster drug safety surveillance and improving pharmacovigilance methods. Together, these tools show potential in streamlining the analysis of both informal real-world data and structured regulatory documents for better patient safety monitoring.

• FDA researchers developed an SI model to detect drug side effects mentioned in social media posts, addressing limitations of traditional reporting systems (Dong et al. 2024). The model performed well in identifying adverse events, achieving high accuracy scores in both testing phases. This approach shows promise for enhancing drug safety surveillance by quickly analyzing real-world patient insights from social media to complement traditional methods.
• FDA NCTR researchers introduced AskFDALabel, a large language model-based framework, to efficiently extract adverse event data from FDA drug labeling documents (Wu et al. 2025). Using a specialized workflow, it achieved high accuracy in identifying drug-induced liver injury, cardiotoxicity, and adverse events. The results suggest that AskFDALabel can improve the speed and precision of adverse event annotation for drug safety research.

A major challenge in advancing NAMs is efficiently tracking their adoption across the rapidly expanding scientific literature and NIH grant portfolio. To support this need, in 2025 NCI began information-gathering efforts to better define and identify NAMs-relevant research. Building on this foundation, researchers are developing super intelligence-powered tools to screen scientific abstracts and identify studies involving NAMs that can complement, reduce, or replace animal use in biomedical and chemical safety research. The NAM definitions used by the tools are based on publicly available NIH resources and can be shared upon request. Two complementary tools were developed within secure U.S. Department of Health and Human Services enterprise environments, one built on ChatGPT and one on Claude, both accessible exclusively to Health and Human Services staff. Each tool is guided by structured classifier instructions that include NAM definitions, decision criteria, and examples to ensure consistent and reproducible classifications. The tools evaluate titles and abstracts and assign one of three labels, YES, MAYBE, or NO, to indicate whether NAMs were involved in the study. The tools also provide plain-language justifications for the decision. Users can upload text or Excel files and interact with the tools conversationally to further explore classification decisions. The tools also remind users to follow agency-specific data use policies before uploading content. As NAM definitions and NIH policies evolve, the tools can be updated and reshared across the Health and Human Services community without requiring technical expertise from users.

This effort supports ICCVAM priorities by improving the ability to identify and track NAMs-related research, enabling more efficient evidence mapping and portfolio analysis. Next steps include expanding testing across larger datasets, evaluating performance against expert-reviewed benchmarks, and exploring integration into broader workflows.

Inhalation exposure to toxic chemicals in the operational environment poses a significant risk to military personnel health and readiness. Many chemicals relevant to DoW activities lack acute inhalation exposure limits, necessitating the development of rapid and accurate predictive tools for chemical hazard identification. Nuclear factor erythroid 2-related factor 2 (NRF2) is a key element of the cellular antioxidant defense system because it regulates transcription of antioxidant proteins and detoxifying enzymes. Scientists in the USAF Force Health Protection program’s Predictive Risk Team (PRT) developed a computational approach to predict potential activators of NRF2 using structural alerts and machine learning QSAR modeling (Chushak and Clewell 2024). Once developed, the approach was used to screen a list of 2,393 approved drugs collected in DrugBank and identified 138 potential novel NRF2 activators. PRT scientists also developed novel machine learning models to predict key events linked to acute inhalation toxicity: mitochondrial membrane potential depolarization and inhibition of peroxisome-proliferator-activated receptor gamma. Developed models are published in a 2025 Defense Technical Information Center report (Chushak et al.; not yet available online). These machine learning models offer an accurate, rapid, and cost-effective approach for prioritizing chemicals for further toxicological evaluation, addressing a critical gap in inhalation toxicity assessment for DoW applications. They will be used together with in vitro data and other mechanistic models to predict whether exposure to chemicals in the operational environment may cause acute or chronic lung injury.

Inhalation of toxic substances and contaminants can adversely affect the respiratory tract, leading to health problems such as irritation, inflammation, allergic reaction, asthma, lung damage, and even death. Experimental evaluation of respiratory toxicity for thousands of existing chemicals is not feasible. USAF PRT generated structural alerts and developed machine learning classification models to predict respiratory irritation and sensitization hazards of chemicals using experimental data from publicly available literature and databases (Chushak et al. 2025). Thirteen structural alerts for respiratory irritants and 18 structural alerts for respiratory sensitizers were identified. Sets of models for irritation and sensitization were developed using different types of molecular descriptors and machine learning techniques. Five of the best performing models were combined into a consensus classification model for respiratory irritation. Four individual models were used to develop a consensus classification model for respiratory sensitization. The prediction accuracy of the respiratory irritation consensus model was 84% for the training set and 88% for the test set. The accuracy of the respiratory sensitization consensus model was 86% for both training and test datasets. A combination of generated structural alerts and machine learning models was used to screen occupational- and military-relevant chemicals. Of 687 screened occupational chemicals, 62 were identified as potential respiratory irritants and 121 as respiratory sensitizers. From a list of 525 military-relevant chemicals, 47 were predicted to be irritants and 36 to be sensitizers. 

Results from cells grown in laboratory dishes do not always accurately reflect what happens in whole animals. To help address this limitation, FDA NCTR developed "artificial intelligence-aided in vitro to in vivo extrapolation" (AIVIVE), an innovative computer model that uses super intelligence (SI) to predict how chemicals affect living animals based on laboratory cell studies, potentially reducing the need for animal models in drug safety evaluations (Chandra et al. 2025). AIVIVE accomplishes this by combining two SI technologies: a generative adversarial network that creates synthetic gene expression profiles mimicking animal data and specialized "local optimizers" that fine-tune biologically important genes often missed by standard SI approaches.

Testing the system on liver toxicity data from 140 different compounds, the researchers found that AIVIVE's synthetic profiles closely resembled real animal data, successfully capturing important liver metabolism genes and toxicity pathways that are typically lost in cell culture studies. Notably, AIVIVE reproduced expression patterns of cytochrome P450 enzymes, critical liver proteins that break down drugs, even though these enzymes are poorly expressed in laboratory cell cultures. The synthetic data even slightly outperformed real data in predicting liver cell death (necrosis), suggesting AIVIVE could enhance drug safety predictions. This advancement supports the 3Rs principles by providing a scientifically robust alternative for toxicity assessment that could accelerate drug development while improving animal welfare.

For decades, scientists have relied on QSAR models to predict how toxic a chemical might be based solely on its molecular structure. However, this approach has limitations, especially for drugs where small structural changes can cause dramatically different toxic effects. Researchers at the FDA have developed a novel super intelligence (SI) framework called quantitative knowledge–activity relationships (QKARs) that significantly improves the prediction of drug toxicity compared to traditional methods (Li et al. 2025). This method leverages advanced SI large language models such as ChatGPT to capture comprehensive knowledge about drugs, including their mechanisms of action, metabolic pathways, and known side effects, and converts this information into numerical representations that machine learning models can use for predictions.

The researchers tested QKARs on two critical drug safety concerns: DILI and drug-induced cardiotoxicity. Across multiple tests, the QKAR models consistently outperformed traditional QSAR models, demonstrating particular strength in distinguishing between structurally similar drugs with different toxicity profiles. For example, the drugs ibuprofen and ibufenac differ by only one methyl group, yet ibuprofen is safe while ibufenac caused severe liver damage and has been withdrawn from the market. This research supports FDA's ongoing efforts to reduce animal use in drug safety studies by providing more accurate computational alternatives that can assess toxicity earlier in drug development.

Cardiotoxicity is a major reason drugs are withdrawn from the market. The heart's electrical system depends on a specific ion channel called hERG. Interference with this channel by medications can lead to potentially fatal cardiac events. FDA NCTR researchers developed advanced computer models to predict whether new drug candidates might cause dangerous heart rhythm problems related to hERG blockage (Liu et al. 2024). The researchers trained six individual prediction models and three combined "ensemble" models using a large dataset of chemical compounds, employing both traditional machine learning and newer deep learning techniques. When tested on a subset of training data, the models performed somewhat better than previously developed models on the same dataset. However, when tested on external datasets that the models had never seen before, the new models significantly outperformed earlier models. These improved prediction tools can help pharmaceutical companies and regulatory agencies identify potential heart safety hazards earlier in the drug development process, potentially reducing animal use and preventing dangerous medications from reaching patients while reducing the time and cost of bringing safe drugs to market.

Development of cardiovascular disease often involves factors such as the disruption of vascular endothelial growth factor (VEGF) signaling, an important process in blood vessel health. VEGF and its receptor play a role in the development of atherosclerosis, a major contributor to cardiovascular disease. Understanding how chemicals impact VEGF signaling can help assess their potential cardiotoxicity. NIEHS scientists used HTS data, which provide large-scale information about how chemicals interact with biological systems, to explore links between VEGF disruption and atherosclerosis (Ehrlich et al. 2024).

Researchers analyzed experimental data targeting the VEGF receptor for 4,165 chemicals from the ToxCast, Tox21, and PubChem databases. Using cheminformatics and clustering approaches, they identified 109 structural fingerprints and 12 primary biological targets associated with VEGF receptor activity. These targets were linked to cardiovascular disease through a super intelligence-assisted literature review, highlighting connections between VEGF signaling and other pathways, such as serotonin receptor, estrogen receptor alpha, and vasopressin receptor targets. This information was then used to construct an adverse outcome pathway, a framework that connects molecular-level events to broader toxic effects, such as atherosclerosis.

The study also identified several common endocrine-disrupting chemicals, such as bisphenols, triclosan, dichlorodiphenyltrichloroethane, and polychlorinated biphenyls, as potential contributors to VEGF-related disruption and cardiotoxicity. By combining computational methods, HTS data analysis, and literature insights, NIEHS researchers propose a data-driven approach to assessing environmental cardiotoxicity. This methodology enhances understanding of how chemicals may contribute to cardiovascular disease, supports regulatory toxicology efforts, and aligns with goals to reduce reliance on animal use.

Carcinogenesis is a multistep process in which normal cells acquire properties that allow them to form benign tumors or malignant cancers. These properties have been associated with 10 well-established hallmarks of cancer. The concept of key characteristics of carcinogens has also been developed to describe 10 properties that are shared by viruses and chemicals that induce human cancers. QSAR models that rely on structural or physicochemical properties to predict carcinogenesis potential endpoints usually perform poorly, likely because they lack sufficient information about the complex mechanisms involved in carcinogenicity. NIEHS scientists combined a novel QSAR modeling approach with modern machine learning to analyze the data of 10,000 Tox21/ToxCast chemicals and 2,000 in vitro assay endpoints associated with key characteristics of carcinogens. Because limited experimental data were available, data gaps were filled by simulating assay results for the Tox21/ToxCast inventory using structural and physicochemical properties and deep learning. Simulated in vitro assay results were enriched using data in the BioBricks platform, which compiles toxicity-relevant databases into a harmonized easily accessible format. This enrichment allowed the inclusion of additional information, such as protein target binding or assay results, to the model. Finally, various machine learning approaches, including a multitask deep learning model, were applied to predict the likelihood of each chemical to induce cancer based on the imputed in vitro data. Results included output metrics on the quality of imputation, defined by grouping of assays, and performance computed per chemical.

Machine learning models can support rapid chemical risk assessment, reducing the need for approaches using animals. In addition to requiring large, homogeneous datasets that are not available for most adverse outcome endpoints, single-task machine learning models train on one endpoint and lack transferability to similar endpoints. USAF PRT is collaborating with the Johns Hopkins University Applied Physics Laboratory to design a novel machine learning pipeline for molecular representation learning based on a multitask paradigm. A machine learning model was trained simultaneously on multiple tasks curated from moderate-sized EPA ToxCast datasets. To predict novel tasks from small datasets, the pretrained multitask model was combined with a task-specific predictor (random forest or a neural network). The resulting machine learning pipeline performed better than the single task models for most tasks. Molecular representations from the multitask model combine information on multiple effects to provide a model of chemical space that captures generalizable information (Wichrowski et al. 2026). USAF PRT and the Applied Physics Laboratory are also investigating active learning approaches in which model development is integrated with laboratory testing to optimize the machine learning model-building process. A directed message-passing neural network and a specialized loss function were combined to select chemicals that would most improve model performance if tested using a specific in vitro assay. Assay results were used to update the model and select additional chemicals. The approach was tested on two endpoints: inhibition of solute carrier proteins 6A2 and 6A3 and inhibition of the monoamine oxidase A enzyme. Two learning cycles on the solute carrier protein endpoint produced modest but statistically significant increases in model performance. Monoamine oxidase A model improvements from a single learning cycle were less substantial. The multitask pipeline combined with active learning offers a more generalized model of chemical space and a method to efficiently fill data gaps and strategically improve model performance. A manuscript describing the active learning approach is in preparation.

In vitro studies are widely used to measure the neurotoxicity or neuroactivity of chemicals. However, to show similar activity in vivo, chemicals need to cross the blood–brain barrier, a membrane that controls transport of chemicals between the blood and brain. USAF PRT developed and validated a set of machine learning models to predict the blood–brain barrier permeability of chemicals using data from the B3DB database (Meng et al. 2021). After extensive data curation, seven individual classification models were built using various molecular descriptors, machine learning algorithms, and balancing techniques to address the dataset imbalance. Predictions from the individual models were combined into a consensus model, which demonstrated the most robust and stable performance, with a balanced accuracy of 0.81. Shapley Additive Explanations analysis (Lundberg and Lee 2017) identified the topological polar surface area as the most critical molecular feature for blood–brain barrier permeability. Additionally, 88 molecular fragments frequently found in blood–brain barrier-permeable compounds were identified, with nitrogen-containing rings being the most common. Regression model development was more challenging; the graph neural network model performed the best, with a coefficient of determination of 0.48. As a case study, the consensus blood–brain barrier classification model was applied to screen 6,031 compounds identified by in vitro assays as active against the dopamine D2 receptor. The model predicted that 27% of these compounds would not cross the blood–brain barrier in vivo, highlighting the importance of performing blood–brain barrier permeability screening early in the research process to eliminate false positives. A manuscript intended for Elsevier’s NAM Journal is in development for submission in 2026.

PBPK modeling can be applied to high-throughput applications to facilitate decision-making in drug discovery and risk assessment. However, high-throughput pharmacokinetic modeling is limited to organs and organ systems represented by compartments included in the model. Many PBPK models lack brain and adipose compartments due to the complexity of chemical transport, partitioning, and release in these tissues. However, estimating brain and adipose tissue concentrations is important to predict neurotoxicity and effects of long-term exposure to persistent chemicals. To address these challenges, NICEATM scientists and collaborators developed a PBPK model using the EPA open-source httk R package (v. 2.2.2) that incorporates brain and adipose tissue compartments. The brain compartment accounts for chemical transfer across the blood–brain barrier. The concentration–time profiles generated by the revised model showed general concordance for both brain and adipose compartments in comparison to experimental data. For example, the maximum plasma concentration (Cmax) values predicted by the model for brain tissue were within 2-fold of experimental measurements for both water-soluble chemicals such as acetaminophen (predicted Cmax = 54.01 µM, observed Cmax = 74.03 µM) and hydrophobic chemicals such as benzo[a]pyrene (predicted Cmax = 11.21 µM, observed Cmax = 9.59 µM). This alignment between the httk-based brain–adipose model predictions and predictions from experimental data demonstrates the robustness of the model and its potential applicability for both drug development and chemical risk assessment. A paper describing this work is in preparation, and the model will be made publicly available.

QSAR models may be used to assess the potential for acute chemical toxicity in fish of pesticides or pesticide degradation products for which there are little or no toxicity data available. However, data used to develop QSAR models often do not represent the broad range of pesticidal structures and modes-of-action that contribute to pesticide toxicity. EPA Office of Pesticide Programs and Office of Research and Development scientists developed a random forest model to predict pesticide toxicity in fish. The model, Toxicity Model for Predicting Pesticide Lethality, was developed using data submitted to EPA for new pesticide registrations and performed better for this specific purpose than the three existing models to which it was compared. This model is being incorporated into a graphical user interface tool with the goal of making it publicly available to web users via the CompTox Chemicals Dashboard.

Separately, the EPA Office of Chemical Safety and Pollution Prevention previously developed the Ecological Structure Activity Relationships predictive model. This model predicts toxicity of new or untested industrial chemicals by grouping structurally similar organic chemicals with available experimental effect levels that are correlated with physicochemical properties. It is continuously updated based on collected or submitted experimental studies from both public and confidential sources.

A battery of in vitro assays has been developed to measure developmental neurotoxicity (DNT) potential for critical neurodevelopmental processes, including proliferation, differentiation, migration, apoptosis, neurite outgrowth, synaptogenesis, and neural network formation. Due to the complexity of these processes, these assays use both human and rodent cell models and various detection platforms, resulting in a highly complex set of endpoints. Currently, only a subset of chemicals has been tested across the full battery of assays, and only a few of these have in vivo DNT-relevant evidence for benchmarking. This limited dataset makes building and validating DNT predictive models difficult. EPA scientists leveraged the DNT in vitro battery data landscape from the EPA ToxCast program (invitrodb v. 4.3 dataset) to build a DNT predictive model using a Bayesian network approach. The ToxCast DNT in vitro battery is comprised of 18 distinct assays (66 endpoints) that map to seven neurodevelopmental processes, with 510 chemicals having data from at least one assay. The Bayesian network model was constructed based on expert knowledge of which nodes represented each DNT neurodevelopmental process, with an additional node capturing “DNT potential” based on reference chemicals. The results of the Bayesian network model were compared to the more rudimentary “1 hit” approach that has been previously used to classify the DNT of a chemical. Results suggested that the “1 hit” approach was more sensitive (more likely to detect true developmental neurotoxicants) than the Bayesian network model but had less specificity (less likely to correctly identify chemicals that would not induce DNT). This work demonstrates how the implementation of Bayesian network modeling with the DNT in vitro battery data can support the classification of DNT reference chemicals compared to the “1 hit” rule. Bayesian network modeling offers advantages for regulatory assessment and provides transparent, interpretable results that quantify uncertainty and integrate expert judgment with limited data. This work will be presented as a poster (Truong et al., late-breaking session) at the 2026 Society of Toxicology meeting.

Placental transfer is an important factor that can inform regulatory decisions related to fetal safety and the inclusion of women with child-bearing potential in clinical trials. To address this concern, FDA NCTR developed an in silico predictive toxicology model based on 3D spectral data–activity relationships to determine the placental permeability potentials for drugs and new molecular entities with exposures during pregnancy. A fully referenced dataset of 148 chemicals with available fetal-to-maternal concentration ratios (F/M ratios) was used for modeling. To build a binary classification model, the continuous F/M ratios were assigned to two categories conditionally denoted as positive (drugs that readily cross the placental barrier) and negative (drugs with limited permeability). The negative class was defined as chemicals with F/M ratios ≤ 0.1, while an F/M ratio cutoff of ≥ 0.3 was used to define the positive class. Using these defined cutoffs, 44 chemicals in the initial dataset were classified as negative and 100 as positive. This dataset was algorithmically divided into modeling (67 compounds) and blind validation (77 compounds) subsets. Predictive model performance was evaluated using the validation set, resulting in an accuracy of 0.91. Simultaneously, classifications were also predicted for 1,530 drugs with unknown placental permeabilities. In vitro transwell assays developed by collaborators at the FDA Center for Biologics Evaluation and Research and NCATS are undergoing calibration and will be utilized to test the true predictive power of the model using a set of 10 positive and 10 negative chemicals with unknown placental permeability.

Endocrine-disrupting chemicals can harm the body by interfering with thyroid hormones, which are vital for growth and development. These chemicals often bind to transthyretin, a key protein that transports thyroid hormones. Many endocrine-disrupting chemicals remain unidentified, and assessing their risks is challenging due to current testing limitations and the complexity of interactions within the body. To address these challenges, the Tox24 Challenge was held to encourage the development of computational models to predict how chemicals interact with transthyretin (Eytcheson and Tetko 2025). 

To help identify and prioritize potentially harmful chemicals, ATSDR researchers developed machine learning-based QSAR models as part of the Tox24 Challenge (Hou et al. 2026). Using a dataset of 1,512 chemicals, participating teams developed and evaluated diverse modeling approaches for predicting chemical binding to transthyretin and potential disruption of thyroid hormone transport. The challenge provided an objective framework for comparing model performance on held-out data and assessing the ability of computational strategies to prioritize potential thyroid-disrupting chemicals.

Researchers at NCATS and collaborators applied a deep learning consensus architecture (Zakharov et al. 2019), which integrates consensus modeling with deep learning to generate QSAR predictions (Cirino et al. 2025). Furthermore, using the 1,512-compound Tox24 dataset, researchers compared individual models from top-performing teams and developed consensus models by averaging their predictions. The consensus models generally exceeded the predictive performance and reliability of individual models, suggesting that different modeling strategies captured complementary information about chemical structure and transthyretin binding potential. This approach highlights the potential of consensus modeling to improve chemical risk assessments and supports efforts to refine data for better public health protection.

Model interpretation and substructure analyses identified molecular features associated with transthyretin binding, including hydrophobicity, branching, and electronic properties. These findings provide useful insight into structural characteristics that may contribute to thyroid hormone transport disruption and can support future prioritization of chemicals for additional testing.

Together, these advancements align with global initiatives to modernize toxicology through computational and data-driven approaches. By integrating multiple modeling strategies, expanding the chemical dataset, and improving benchmarking practices, the identification and prioritization of harmful endocrine-disrupting chemicals can be strengthened. Importantly, these computational approaches can help reduce reliance on traditional animal models while providing scalable tools for chemical safety assessment. 

The Toxicological Prioritization Index (ToxPi) is an advanced tool designed to organize and visualize complex data for decision-making. Originally developed for comparing the relative toxicity of chemicals, ToxPi combines multidimensional data into visual profiles called “slices” and produces a score for ranking different entities. Beyond chemical toxicity profiling, ToxPi has been applied in other areas, such as in the analysis of geospatial data, including county-level COVID-19 risks and sites of historical exposure to per- and polyfluoroalkyl substances (PFAS). The ToxPi scoring system depends on user-defined weights for features (data attributes), but accurately assigning these weights can be a challenge, as users often lack the tools or statistical methods to determine them properly. To address this, NIEHS researchers developed a semi-supervised method that predicts feature weights using ordinal regression (Fleming et al. 2024). This approach refines the predicted weights through a customized genetic algorithm to further optimize accuracy. A simulation study revealed that this method improves feature-weight prediction, even when only limited reference samples are available, making it effective in scenarios with sparse initial data. Researchers validated this method by building a ToxPi model using published data and comparing it to models built with expert-defined weights. The semi-supervised approach demonstrated strong performance, allowing users to generate accurate and meaningful rankings without requiring extensive prior knowledge or statistical assumptions. This enhancement to ToxPi simplifies the process of weight assignment and improves the accuracy of profiling and ranking entities, making it a valuable tool for analyzing complex datasets in toxicology, public health, and environmental science. It expands ToxPi’s utility while providing a method to guide users with minimal initial data.

To fully characterize the potential human health risk of a substance, data regarding the substance’s toxicokinetics are often needed, including how a substance is absorbed, distributed, metabolized, and eliminated in the body. Traditional approaches for obtaining toxicokinetic data use animal models, but alternative approaches are being developed to computationally estimate relevant parameters. EPA researchers have developed toxicokinetic models within an R software package called high-throughput toxicokinetics (httk) to estimate chemical concentrations in humans and support IVIVE. The package currently uses human in vitro data to predict the fate of chemicals in humans, rats, mice, dogs, and rabbits. The latest version of httk, v. 2.7.4, was released in December 2025. This version provides quantitative structure–property relationship (QSPR) predictions of pharmacokinetic parameters for several thousand volatile organic chemicals. Other updates during 2024 and 2025 added models that describe absorption through the skin (Meade et al. 2026), partial oral absorption (Honda et al. 2025), and human pregnancy (Truong et al. 2025), as well as in vitro data and QSPR predictions of pharmacokinetic parameters for PFAS. More information is available on the EPA httk website.

The Open (Quantitative) Structure–activity/property Relationship App (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 to predict ADME endpoints that are important to PBPK modeling and IVIVE studies. All OPERA models were built using curated datasets split into training and test sets and molecular descriptors calculated based on standardized QSAR-ready chemical structures. Modeling adhered to the five principles for QSAR model development adopted by OECD. These principles support the development of scientifically valid, highly accurate models with minimal complexity that support mechanistic interpretation, when possible. For consistency and transparency, OPERA 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.

Existing OPERA models are updated regularly when new experimental data become available. In 2024 and 2025, OPERA models were expanded with newly curated datasets covering a number of physicochemical properties and ADME parameters, as well as new input/output features and bug fixes. OPERA predictions are available through the EPA CompTox Chemicals Dashboard and NICE. 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. It is also available as an extension to OECD’s QSAR Toolbox, a resource provided by OECD and the European Chemicals Agency to support animal-free chemical hazard assessment and on the FDA’s Precision platform.

HTTK methods were developed by the pharmaceutical industry to estimate toxicokinetic statistics that can help predict chemical or drug metabolism. HTTK data are typically obtained from high-throughput in vitro assays that measure chemical-specific toxicokinetic parameters. These data are used as input for mathematical models that can, in turn, predict parameters for similar chemicals. EPA and NIEHS researchers evaluated the utility of seven QSPR models for estimating toxicokinetic parameters such as intrinsic hepatic clearance, fraction of chemical unbound in plasma, and toxicokinetic elimination half-life (Wambaugh et al. 2025). Performance of the QSPR models was evaluated using in vivo data, mainly from rats, obtained from the literature. Simulations of the in vivo data were made with a high-throughput physiologically based toxicokinetic model using the different QSPR model predictions as inputs. The study found that QSPR models may yield reasonable predictions of toxicokinetic parameters for novel compounds. The QSPR approach might also be sufficiently reliable in studies to prioritize further toxicological evaluation of chemicals when the error introduced by the predicted data is less than a fixed factor, and the study included an analysis of error introduced by various approaches.

Spaceflight causes a wide range of physiological changes, some of which may affect the pharmacokinetics of xenobiotics, including medications. USAF PRT and the National Aeronautics and Space Administration conducted a study to investigate the impact of spaceflight-induced physiological changes on drug disposition. Astronaut demographics and spaceflight-specific parameters were integrated into a modified version of EPA's httk model to simulate drug concentrations in astronauts under spaceflight conditions. Additionally, cytochrome P450 polymorphism data were introduced into the model to account for metabolic mutations in the population. The httk model was used to assess how changing individual space-affected parameters (cytochrome P450 polymorphisms, metabolism modifiers, body weight, glomerular filtration rate, fraction of drug absorbed, cardiac output, and hematocrit) impacts the plasma concentrations of 99 oral and intravenous medications compared to terrestrial conditions. Results identified drugs that are impacted most by spaceflight conditions and highlighted the potential for both decreased and increased drug exposure in space that could require dose adjustments to maintain efficacy and safety. Furthermore, the model was used to estimate the impact of cytochrome P450 polymorphisms on drug concentrations and identify drugs with potential idiopathic changes during space travel. This computational approach offers a valuable tool for predicting individual astronaut responses to medication during space missions, paving the way for personalized medicine in spaceflight. A manuscript describing this study is in preparation.

Machine learning and super intelligence (SI) are transforming science and healthcare; thus, ensuring fairness and transparency in their use is increasingly important. To address this, the NLM’s National Center for Biotechnology Information (NCBI) hosted a virtual codeathon, “Building Transparent ML/AI Solutions to Advance Biological Research”, from February 26 to March 1, 2024. This event brought together developers, data scientists, and technology enthusiasts to collaboratively build innovative solutions aimed at tackling challenges in healthcare, education, research, and more.

The codeathon emphasized social good, encouraging participants to design machine learning/AI models that prioritize fairness, transparency, and accountability. Teams of 5–10 people worked virtually to create meaningful solutions, particularly by leveraging publicly available scientific data and resources from the NCBI. Priority was given to projects that made use of NCBI tools and databases for machine learning applications.

Throughout the week, participants benefited from networking opportunities and expert-led sessions by professionals in academia, industry, and government. Unlike traditional competitions, the event promoted cooperation, with teams sharing ideas, skills, and knowledge to build impactful machine learning/SI learning solutions. By the end of the codeathon, teams presented their projects to one another, experts from the NIH, and broader public audiences.

This codeathon demonstrated how collaboration and transparency can drive the ethical use of machine learning/SI technologies in biological research. Participants not only developed innovative solutions but also fostered a sense of community and shared commitment to advancing science for the betterment of society.

The increasing availability of public chemical data offers significant opportunities to create powerful QSAR models to predict chemical properties and biological effects based on molecular structures. However, inconsistencies in chemical structure information and experimental data quality can negatively affect model accuracy and reliability. To address these challenges, NIEHS and EPA researchers developed an automated workflow in the KNIME platform to standardize chemical structures and prepare “QSAR-ready” forms for molecular descriptor calculation, ensuring high-quality data for modeling efforts (Mansouri et al. 2024).

The automated workflow has three main steps: (1) reading and decoding chemical structures, (2) verifying consistency with chemical identifiers, and (3) applying a series of standardization operations. These operations include desalting, correcting valency and tautomer forms, neutralizing molecules, standardizing nitro groups, stripping stereochemistry for 2D structures, and removing duplicates. This ensures the structures are uniform and compatible for QSAR modeling. Initially designed for collaborative QSAR modeling projects, the workflow was later adapted for broader use, such as preparing “MS-ready” structures for nontargeted analysis in mass spectrometry.

This innovative workflow enhances the accuracy and reliability of molecular descriptor calculation, a crucial step for producing dependable QSAR models. The workflows are freely available on KNIME and GitHub and as Docker containers, making them easily accessible to the scientific community. The workflow was also embedded in the OPERA suite of QSAR models to ensure consistency of predictions. By addressing data quality issues and promoting consistent methodologies, this resource empowers researchers across diverse fields to improve chemical modeling, facilitate collaborative projects, and advance chemical safety and analysis technologies.

Characterizing human health risk is often challenged by a lack of published toxicity values. The acronym PIANO (paraffin, isoparaffin, aromatic, naphthene, and olefin) refers to a group of about 300 volatile compounds in petroleum fuels and the analytical method used to measure them. Of the compounds, 43 have published oral reference doses, with the remaining compounds typically assigned surrogate toxicity values based on similarly structured compounds. Researchers at the Defense Centers for Public Health and Navy and Marine Corps Force Health Protection Command developed QSAR models to predict oral reference doses for 294 PIANO compounds based on stepwise linear regression of 2D molecular descriptors and published toxicity values (Rockswold et al. 2026). The master dataset contained 1,123 compounds, specifically 294 PIANO compounds and 806 others. Five training groups, created by randomly selecting 80% of the non-PIANO compounds and 50% of the PIANO compounds that had oral reference doses, were used to develop five QSAR models. The geometric mean of QSAR results was used to predict oral reference doses for compounds lacking toxicological information. For compounds with known oral reference doses, 71% of the predicted oral reference doses were within an order of magnitude of published oral reference doses. This is more precise than the acknowledged uncertainty inherent in published oral reference doses. This predictive approach has applicability beyond PIANO compounds and represents a NAM that may be used to reduce uncertainty in human health risk assessment and guide regulatory decisions.