Before a new drug hits the market, it must undergo years of rigorous preclinical testing and multiple phases of human trials, often involving hundreds or even thousands of patients. This highly regulated process is designed to assess the drug’s toxicity, safety, and efficacy, ultimately safeguarding public health. Even after approval, researchers continue to monitor the drug’s effects to ensure its benefits outweigh any risks.

However, when it comes to the thousands of synthetic chemicals we encounter day to day, it’s a different story. Everyday products, from cosmetics and cleaning supplies to plastics and processed foods, contain chemicals that we know little to nothing about. These chemicals, often designed to enhance product performance or extend shelf life, are increasingly widespread, raising concerns about their long-term effects on ecosystems and human health.¹ In the United States, the Environmental Protection Agency (EPA) oversees the regulation of chemicals produced and distributed for commercial and industrial purposes, provided they do not fall under the jurisdiction of other agencies, such as the Food and Drug Administration (FDA).

In 1976, President Gerald Ford signed the Toxic Substances Control Act (TSCA), which, in theory, gave the recently formed EPA regulatory authority to require toxicity testing and risk assessment reports on chemicals intended for commercial use. However, public health specialists argue that in practice, the law provides the EPA with limited authority to effectively regulate chemicals, especially in terms of pre-market approval, comprehensive safety evaluations, and the ability to restrict or ban substances deemed harmful.^2,3

Moreover, the nearly 60,000 chemical substances already in use were automatically deemed safe and grandfathered into an inventory of existing chemicals, despite little to no evidence supporting their safety. Over the years, the TSCA’s inventory of chemical substances approved for use has grown to more than 80,000, yet safety data on these chemicals remains scant.

“This backlog of testing is simply enormous,” said Thomas Hartung, a toxicologist at Johns Hopkins University. “Nobody can close this testing gap.”

Traditional toxicity reports and long-term epidemiological data, which are critical for assessing risks to human health, are both incredibly expensive and time-consuming to obtain. Testing a single chemical can take months or even years, and the endpoints assessed are often limited in scope. In response to these challenges, government agencies responsible for overseeing toxicology testing came together twenty years ago to envision a new approach to toxicology testing—one better suited to the needs and complexities of the 21^st century.⁴

A New Vision and a Roadmap for Toxicity Testing

The bioinformatics boom of the 1990s, along with subsequent advances in genomics, proteomics, and metabolomics, revolutionized the life sciences. These developments expanded our understanding of human disease and transformed how scientists evaluate the safety and efficacy of pharmaceutical products. Yet, some fields lagged in incorporating these powerful new tools.

“Toxicology wasn’t really adapting and folding those into the testing paradigm,” said Russell Thomas, a bioinformatician at the EPA, where he is director of the Center for Computational Toxicology and Exposure (CCTE). However, just after the turn of the century, things started to change.

In 2004, the National Toxicology Program (NTP), based at the National Institute of Environmental Health Sciences (NIEHS), released its vision and roadmap for the future of toxicology testing.⁵ The report laid out a framework to advance the field of toxicology from an observational science to a predictive science driven by mechanism-based biological data.

The EPA was also in the process of modernizing toxicity testing to meet evolving regulatory needs and address the backlog of substances requiring evaluation. At the EPA’s request, the National Research Council (NRC) conducted a comprehensive review of existing toxicity testing methods, their limitations, and strategies for improvement. This effort culminated in a 2007 report outlining a long-term vision for transforming toxicity testing.⁶ The NRC proposed high-throughput in vitro testing as the primary approach, with in vivo testing, particularly in nonmammalian species, used selectively to fill knowledge gaps.

“ [These reports] were really in response to people from every sector—scientists, regulators, industry groups, community groups, advocacy organizations—who were looking at the buildup of new chemicals that were being put into commerce and the environment and realizing that the way we had set up to do testing, which was invented before there was a concept of this scope of the problem, was never going to keep up,” said David Reif, a geneticist and head of the Predictive Toxicology Branch at the NIEHS.

These reports helped give rise to a unique interagency collaboration, known as Tox21, which formed in 2008. This partnership, which remains active today, brings together the EPA, NIEHS, FDA, and the National Center for Advancing Translational Sciences (NCATS).

The goals of the Tox21 consortium are to uncover key biological pathways that, when disrupted, can cause adverse health outcomes, prioritize chemicals for further testing, and develop more accurate models for predicting in vivo toxicological responses. More than a decade into the program, it has achieved notable successes that hold the potential to transform how chemicals are evaluated for their health effects.

High-content and high-throughput screening have been, and remain, essential for advancing the goals of programs like Tox21. In its early days Tox21 relied on high-throughput screening using more targeted functional assays, such as those measuring chemical binding to specific receptors. However, Thomas highlighted a shift toward broader profiling assays, which offer a more comprehensive view of the potential processes and pathways affected by a chemical. For example, high-throughput transcriptomics and high-content imaging can help scientists comprehensively analyze gene expression and capture detailed cellular responses to chemical exposures on a large scale.

Hundreds of in vitro high-throughput assays are available for generating detailed bioactivity profiles across a wide range of biochemical targets and cellular endpoints. While 96-well plates are standard in many laboratories, the Tox21 program takes plate-based assays to the next level. Using a sophisticated robotic arm housed at NCATS, the program can manage and store just over a thousand 1,536-well plates, or approximately 1.5 million assay wells. It would take more than 15,500 painstakingly pipetted 96-well plates to achieve the same output. This advanced robotic setup enables unparalleled throughput and efficiency in chemical testing.

“Robotized, highly standardized testing of many substances is what the field needs,” said Hartung.

Since its launch in 2008, the program has generated over 120 million data points on approximately 8,500 chemicals—many of which had limited empirical data on their biological effects before these efforts.⁷ Now, the Tox21 library, in combination with the EPA’s ToxCast library, represents the largest source of in vitro bioassay data for environmental chemicals, and it is publicly available.⁸ “These are really well developed, very diverse chemical sets, and still they don’t cover all the space we need,” Reif added, emphasizing the vast chemical landscape that scientists and regulators must navigate.

Success Stories from a Transforming System

During the 1990s, growing concerns arose among scientists about certain chemicals—many found in everyday products—with the potential to disrupt human endocrine systems and affect hormone production, development, and reproduction. Small disturbances in endocrine function, particularly at critical developmental periods, can have lasting effects on human health.⁹ These concerns prompted the EPA to establish the Endocrine Disruptor Screening Program (EDSP) in 1998 with the goal of assessing the effects of chemicals on estrogen, androgen, and thyroid hormone systems. Integrating Tox21 data into their computational models enables EDSP researchers to more quickly predict endocrine activity and prioritize specific chemicals for further testing.

In parallel efforts to determine potential endocrine effects of environmental chemicals, researchers at NIEHS and NCATS used the extensive Tox21 compound collection to investigate environmental chemicals potentially linked to early puberty in girls—a rising global trend.¹⁰ They focused on identifying compounds that activate receptors in the hypothalamus, a region of the brain involved in regulating puberty timing. After running multiple high-throughput screens using human cell lines, they identified several compounds that activate receptors for kisspeptin or gonadotropin-releasing hormone (GnRH), both of which play a role initiating puberty. For example, they determined that musk ambrette, a fragrant chemical found in products like soap, detergent, and perfume, activates the kisspeptin receptor. In follow-up experiments, the researchers found that musk ambrette increased expression of the gene that encodes GnRH in human hypothalamic cells and expanded the GnRH neuronal areas in developing zebrafish larvae.

While the Tox21 program is still on its way to realizing many of its initial goals, one undeniable achievement is the generation of an extensive dataset on chemical characterization. As toxicology testing continues to transition toward high-content, high-throughput in vitro assay approaches, the ability to interpret and make sense of the vast amounts of data will become increasingly critical to the program’s success and its broader impact on the field.

AI: The Crystal Ball That Toxicology Needs?

With the completion of the Human Genome Project in the early 2000s, many anticipated that the mysteries of the human genetic blueprint would finally be unlocked. “The original thought was, now that we have the data, the patterns will emerge,” said Reif. However, the project raised more questions than it answered, and more than two decades later, scientists are still working to fully decipher the genome. “The same is true for toxicology,” he added. While the ever-growing datasets are a valuable resource, scientists are still grappling with how best to manage and analyze this vast amount of information.

While each test or high-throughput assay contributes a small piece to the puzzle of chemical characterization, there is no single test that can provide all the answers. To piece together the full picture, computational tools have played a crucial role. “The analysis part has been actually one of the most challenging parts,” said Reif.

Each federal agency has its own dedicated branch for computational toxicology, where bioinformatics and machine learning approaches are used to interpret data and assess the potential health effects of thousands of chemicals. While artificial intelligence (AI) methods have been integrated into analysis workflows for years, recent advances in AI are poised to revolutionize toxicology research. By enabling faster and more accurate predictions of chemical toxicity through advanced data analysis, machine learning models, and simulations, AI has the potential to accelerate the evaluation of substances and reduce the reliance on animal testing.

“AI is going to really revolutionize a lot of things,” said Reif.

“AI only shines when there is big data, and high-throughput screening is one way of mass-producing data which characterize chemicals,” said Hartung. He emphasized that Tox21 is becoming increasingly valuable with recent advances in AI because it offers an enormous repository of chemical characterization data.

A long-term goal in toxicology is to forecast the functional characteristics and potential toxic effects of chemicals while reducing the need for extensive testing. “The predictive part—that’s the pinnacle,” said Reif. However, he emphasized that predictive toxicology is dependent on high-quality data. “That’s where I think things like Tox21 and the data generation component will continue to have huge value.”

The Times They Are A-Changin’

In June 2016, the nation’s primary chemicals management law—TSCA—was amended to increase the EPA’s authority to take actions to protect humans and the environment from hazardous chemicals. The Lautenberg Chemical Safety Act, which received bipartisan support from both chambers of Congress, introduced several significant provisions. One key change requires the EPA to evaluate the safety of both existing and new chemicals and prioritize them for comprehensive safety assessments. They must also provide risk-based chemical assessments to the public that guide risk management actions, including bans or use restrictions.

The first 10 chemicals that the EPA prioritized for risk evaluation included the solvents trichloroethylene (TCE) and perchloroethylene (PCE)—volatile organic compounds used in consumer products like cleaning supplies, brake cleaners, lubricants, and adhesives and in dry cleaning. Both chemicals are known to cause liver and kidney cancers and damage the brain and reproductive organs. In December 2024, following a final risk evaluation, the EPA cracked down on most uses of PCE and TCE, with a ban on the substances taking effect within one to three years. However, much work remains ahead. TCE and PCE are just two of nearly 40 chemicals in the TSCA risk evaluations pipeline, and only the tip of the iceberg of existing chemicals that need risk evaluation.

The EPA is also responsible for reviewing new chemicals and significant new uses of existing chemicals. When a company plans to import or manufacture a chemical in the US above a certain tonnage limit, it must submit a pre-manufacture notice to the EPA. Under the 2016 TSCA amendment, the EPA must affirmatively determine whether the chemical poses an unreasonable risk to human health or the environment based on its intended use. Each year, the EPA processes over a thousand submissions, with decisions required within 90 days of receipt.

“Companies typically don’t provide a lot of data associated with that pre-manufacturing notice,” said Thomas. “It’s usually in very data-poor space.”

To address the gaps in chemical safety information in new chemical applications, the EPA is adopting novel approaches to modernize how the agency evaluates these chemicals. These initiatives include using computational models to predict potential risks of new chemicals based on data from structurally similar substances and developing a tool that integrates diverse data streams, such as information from the Tox21 chemical library.

“That’s a new program that I think is particularly exciting and shows the maturing of these Tox21 type of initiatives to have an impact, not just on some of those endocrine programs, but also the evaluation of new chemicals as they enter the US economy,” said Thomas.

But keeping these programs running requires money, and lots of it given the sheer number of chemicals that still need to undergo thorough evaluations. Many fear that a second Donald Trump presidency will act on previous promises to slash funding for the EPA and loosen environmental regulations. The EPA’s integrity may also be at risk: During the first Trump administration, a political appointee meddled with the agency’s assessment of a hazardous per- and polyfluoroalkyl (PFAS) compound, weakening the assessment just before it was set to be published.

While the development of new technologies and the collection of chemical toxicity data are critical, this must be accompanied by policies that ensure sustained funding for these efforts and grant federal agencies the authority to fulfill their duty to protect human health and the environment.