Is This the End of Animal Testing? FDA Announces Plans to Phase Out Animals in Drug Safety Studies

For decades, preclinical testing in mice, rats, and nonhuman primates has been a crucial part of drug development, an important although not infallible way to ensure some measure of safety for human participants in clinical trials. In 2022, Congress passed the FDA Modernization Act 2.0, stating that the agency was no longer compelled by law to require animal testing.¹ However, the act did not prohibit the FDA from requiring animal testing, and animal toxicity data remained an essential step in the path towards human trials.

But on April 10, the FDA announced plans to phase out these requirements, stating that they would be “reduced, refined, or potentially replaced” with New Approach Methodologies (NAMs). These methods include human-derived cell models, such as organoids and organ-on-a-chip systems, as well as in silico approaches such as pharmacokinetic modeling and toxicity-predicting machine learning algorithms. Within the year, certain monoclonal antibodies, which are most commonly used in the treatment of cancer and serious autoimmune disease, could be evaluated using a “primarily non-animal-based testing strategy.” According to the roadmap accompanying the FDA announcement, “In the long-term (3–5 years), FDA will aim to make animal studies the exception rather than the norm for pre-clinical safety/toxicity testing.”

This announcement has been met with both hopefulness and concern in the scientific community. On one hand, there is broad support for moving away from animal testing, which is ethically fraught, expensive, and not always predictive of human biological responses. On the other hand, scientists and pharmaceutical industry professionals have expressed that NAMs are not yet advanced enough to fully replace animal models.^2,3 Moreover, there are concerns that dramatic reductions in budgets for scientific research and the firing of thousands of workers at the FDA and NIH will stall further development of NAMs and interfere with the functioning of the very systems that would be responsible for validating, standardizing, and monitoring the efficacy of these technologies.

Alex Rubinsteyn, a University of North Carolina at Chapel Hill researcher who uses machine learning approaches to inform the development of personalized cancer vaccines, supports reducing animal testing but is uncertain about how this will play out in practice in the current climate. “I think this could become a disaster,” he said. “But it could also potentially unlock a much faster rate of progress.”

Few would dispute the shortcomings of current animal testing pathways: About 90 percent of drugs that make it through preclinical trials never obtain FDA approval for use in humans, largely due to insufficient efficacy or safety.⁴ Joseph Wu, who studies patient-specific in vitro models of cardiovascular disease at Stanford University, said that this high failure rate is partly due to the inherent differences between human and rodent biology. Furthermore, he noted, “The heterogeneity that exists among humans cannot be captured by using a traditional mouse model.”

Additionally, despite attempts to streamline evaluations of drugs for currently untreatable diseases, “[Drug development] is still really slow and really expensive,” said Rubinsteyn. “And it's unmatched to clinical realities for certain kinds of disease: There are sufficiently deadly diseases where you really would want to go much faster than you're allowed to go.”

How Do In Vitro and In Silico Approaches Stack Up Against In Vivo?

Despite the inherent flaws in animal testing, many researchers say that NAMs are not yet advanced enough to fully replace traditional drug safety studies. For example, in a late 2024 report to the FDA Science Board, members of the NAMs subcommittee—convened in 2023 to provide recommendations on integrating NAMs into regulatory processes—wrote that, “Technical limitations to current NAMs exist…today, no assays fully capture the critical hazard endpoints for assessing all currently existing human or animal organ systems; therefore, NAMs cannot fully eliminate the use of integrated physiological systems such as in animal and human trials.”²

In a published response to the FDA announcement, the president of the National Association for Biomedical Research, Matthew Bailey, echoed these sentiments. “No AI model or simulation has yet demonstrated the ability to fully replicate all the unknowns about many full biological systems.”

Similarly, Rubinsteyn noted that while these models can be useful when the space is sufficiently constrained, in other situations, they are still no match for the complexity of biology. “The thing that I work on the most—personalized cancer vaccines—is totally plagued by machine learning models not capturing the relevant realm.”

“We could improve those models with cell lines, but ultimately, the cell lines will be different if we put them in the context of a living organism,” he continued. “[In vivo], the tumor cells will shift what they express. They have to deal with being in contact with other cells in the tissue. They'll have to pull in vasculature, and they'll have to deal with the immune environment. So, they're going to shift how they behave.”

To address this problem, other research groups are building organ-on-a-chip systems as a closer approximation of how cells behave in living tissues. These models can have layers of cells supported by an extracellular matrix with a simplified vascular system, recapitulating some of the cell-cell interactions and mechanical forces present in a living organism.

Yu Shrike Zhang, a Harvard Medical School researcher who uses bioprinting, microfluidics, and other techniques to create improved organ-on-a-chip platforms, noted that these models are quite advanced for certain tissue types.⁵ “For the liver, the models are pretty precise in general,” Zhang said. “It's been [studied] for a very long time, and people know exactly how it works.”

Indeed, a liver-on-a-chip model developed by the biotechnology company Emulate, Inc., was able to correctly identify drugs known to be toxic or nontoxic to the liver with a sensitivity of 87 percent and a specificity of 100 percent.⁶

Using organ-on-a-chip models, Zhang said, “In three to five years, I think we can probably get to a pretty high level in terms of testing [toxicity or biological responses] for individual organs.” Modeling interactions between different organs, however, is a more challenging task. Crosstalk between organ systems is complex and incompletely understood, and organs can be influenced, or influence each other, via changes in metabolism, blood circulation, immune function, endocrine signalling, or nervous system activity.⁷

Scaling is also a concern, according to Zhang. As size changes, different physical forces and properties increase or decrease in different ways. So, it is not yet entirely clear how to best model a 200-pound human body using organ-on-a-chip systems, some of which may be only a few cell layers thick.

“Things are quite complicated, both biologically and in terms of how these devices operate,” said Zhang. “Being able to really reproduce that organism-level interaction—I think that's still something that will be really important to look into. Maybe that's something that's going to be mature in the next three to five years? I mean, no one knows. But I think that's probably one of the major limitations right now.”

Refine, Reduce, Replace and the Promise of NAMS

Even scientists who are extremely enthusiastic about NAMs seem to view them as a tool to reduce, not completely replace animal testing. Wu, for example, has spent decades developing in vitro models using cardiomyocytes derived from human induced pluripotent stem cells (iPSCs) to improve our understanding of cardiovascular disease. He has created an extensive biobank of human iPSCs, capturing genetic diversity in health and disease, and even founded a company, Greenstone Biosciences, which aims to accelerate the drug discovery process by combining in vitro and in silico approaches.

However, Wu said, “I'm a proponent of using all models. I'm not a proponent of saying that, ‘Oh, in the future, we should just get rid of mouse models.’” Instead, he said, applying these strategies prior to animal testing could greatly reduce the number of animals that would be needed for each experiment, enabling researchers to identify promising targets and screen for well-defined types of toxicity.

For example, Wu was part of a project led by fellow Stanford University cardiovascular biologist Mark Mercola, in which the team used iPSC-derived human cardiomyocytes and machine learning to classify existing drugs as low risk or intermediate/high risk for causing dangerous arrhythmias. Using area under the curve as a measure of the model’s accuracy—for which 0.5 indicates a random classifier and one is a perfect classifier—the model correctly identified risk with an area under the curve value of 0.95.⁸ In the future, a system like this one could help researchers spot potentially cardiotoxic compounds early in the drug development process. This has the potential to prevent the investment of time, money, and animal lives into investigating a drug that might treat one disease very well but be ultimately useless because of severe adverse effects.

Furthermore, unlike studies performed in strains of genetically identical mice, research with human cells can provide not only general safety predictions, but they also help identify which individuals might be most at risk for particular side effects and even suggest mechanisms for mitigating these effects.

For example, the chemotherapy drug doxorubicin can lead to heart failure in a subset of patients, but for many years, the mechanism of this cardiotoxicity was not known, and there was no way to predict which patients were at risk. In a study of eight breast cancer patients, Wu and his team showed that iPSC-derived cardiomyocytes from patients who experienced this side effect were more sensitive to doxorubicin toxicity than cells from patients who did not. ⁹ In the future, this could serve as a tool for screening patients prior to treatment. In a subsequent study, the researchers used a CRISPR-based approach to screen cardiomyocytes for genes that contributed to this vulnerability. One gene, which coded for the enzyme carbonic anhydrase 12, seemed to play a large role: When expression of this gene was inhibited in the cells, they were protected from doxorubicin toxicity.¹⁰ An antagonist of this enzyme, Indisulam, was also protective in heart cells. Only after all these experiments did the researchers test the drug in mice.

Since then, Wu and his team have used iPSC-derived cells, patient data, and AI to identify a candidate compound for the treatment of marijuana-induced vasculature inflammation, and two potential therapies for cardiac fibrosis.^11–13 “These three papers all have mouse models, but they're toward the end,” said Wu. “They're only done for validation—the initial screen, initial validation, initial design, all that stuff is done [using] organoids, stem cells, and AI.”

The first candidate is currently in a Phase 1 clinical trial for the treatment of inflammation associated with heart failure, the second is in an open-label study for treating idiopathic pulmonary fibrosis. In the coming years, studies such as these will provide crucial data to answer the question of whether these newer drug development techniques can increase efficiency and reduce failure rates in clinical trials.

An Uncertain Future for Drug Development

Much work remains to be done, however, if animal testing is to be truly replaced in the next three to five years. In addition to the development of the NAMs technologies themselves, the FDA roadmap also calls for the creation of open-access toxicity information databases, developing strategies to validate NAMs, determining appropriate thresholds for eliminating animal testing, figuring out how to standardize these techniques so that they can be compared across many different laboratories, coordinating with other federal agencies, and monitoring how well all of this is working.

“Transitioning from animal-based testing to NAMs for safety will require careful planning, robust science, and collaboration,” the roadmap states.

But will this be possible in the chaos currently afflicting many government agencies and the dramatic changes to support for scientific research in the United States? The Trump administration has already terminated 1.8 billion dollars in National Institutes of Health (NIH) grants; the administration’s proposal for the upcoming year would slash the budget of the NIH by 40 percent.^14,15

Some of these governmental budget cuts and funding freezes adversely impact the laboratories that have been instrumental in developing the very NAMs technologies the FDA is hoping to promote. For example, Harvard University bioengineer Donald Ingber, a pioneer in organ-chip research and scientific founder of Emulate, Inc., received stop-work orders on two major organ-on-a-chip projects in late April 2025.

Beyond the technologies themselves, planning and collaboration efforts may also be impacted by the major changes at these agencies. So far, 2025 has been marked by many cancelled or postponed scientific meetings at the FDA and NIH, as well as firings of thousands of workers, including many top-level officials and a large portion of communications roles, and the resignation of Peter Marks, director of Center for Biologics Evaluation and Research.

Rubinsteyn, for his part, worries about how reductions in animal testing requirements will play out in such an environment, raising concerns that insufficient oversight could create opportunities for unscrupulous companies to bring potentially unsafe drugs to market.

“I do think that this is, in principle, a positive direction for change,” he said. But depending on how these changes are implemented, “it could go quite wrong.”

Disclosure of conflicts of interest: Yu Shrike Zhang sits on the scientific advisory board and holds options with Xellar Biosystems.

References