On September 21, FDA announced a change to the language of its drug safety regulations. References to animal tests and animal studies would become references to nonclinical tests and nonclinical studies.
The direct final rule, published the following day, recognizes a broader range of methods, including computer modeling. FDA said the change neither prohibits animal studies nor changes the evidence standards for drug development.
That raises a more important question than the terminology: what makes a computer model’s prediction reliable enough to help decide whether people can receive an experimental drug?
The method can generate evidence. Researchers and FDA reviewers must determine what that evidence establishes.
Congress Changed the Law First
The change began with Congress. Section 3209 of the Food and Drug Omnibus Reform Act of 2022, enacted within the Consolidated Appropriations Act, 2023, amended the statute to recognize nonclinical tests, expressly including computer modeling, cell-based assays, organ chips, and animal tests.
FDA’s rule brings its regulations into alignment with that broader language. Some provisions still referred specifically to animal studies even though other approaches were available and FDA already accepted appropriate nonanimal data.
Nonclinical describes the kind of evidence, rather than a requirement to use a particular species or testing platform. The category still includes animal studies. It also extends beyond tests conducted before a clinical trial begins.
The rule is scheduled to take effect February 4, 2027. Comments close December 7, 2026. Significant adverse comments can cause FDA to withdraw the direct final rule and proceed through the accompanying proposed rule.
The agency has opened the regulatory language to more methods. Whether a particular method supplies adequate evidence remains a scientific question.
A Different Method Can Reveal a Different Risk
Animals provide information about how a drug affects a living organism. But a result in an animal does not always predict what will happen in a person. Differences in biology can limit the usefulness of an animal model for a particular drug or safety question.
New Approach Methodologies, or NAMs, offer other ways to investigate those questions. A test using human cells can examine a response in human tissue. An organ chip can reproduce selected features of an organ’s environment. A computer model can use data and assumptions about biological processes to estimate a response or exposure.
These methods are broader than AI. Some computer models use AI; others use mathematical representations of biological processes. Cell-based tests and organ chips need not use AI at all.
The reason to consider them is their potential to provide information relevant to human risk. FDA’s March 2026 draft guidance on NAMs describes methods that can improve prediction, fill gaps in existing models, or complement other findings.
A model’s usefulness depends on the question it is asked to answer. Predicting one mechanism of injury does not establish that the model captures every way a drug could cause harm.
FDA Asks What the Method Can Establish
The draft guidance identifies four dimensions for evaluating NAMs. In ordinary language, they ask:
What is the intended use? The method needs a defined question and regulatory purpose, such as informing dose selection or monitoring during a trial.
Why is it relevant to human biology? The developer must explain how the features represented relate to the human risk being assessed.
Does it produce dependable results? Technical characterization examines whether the method is robust, reliable, and reproducible. FDA explicitly connects this expectation to the assessment of animal studies.
Is it adequate for this decision? Fitness for purpose concerns performance, limitations, and how the findings contribute to the overall risk assessment. Where established comparators exist, the draft recommends evaluating whether the NAM characterizes risk at least as well, including false positives and false negatives.
These questions turn an impressive result into an assessable claim. A prediction that misses a relevant toxicity has a different consequence from one that flags a risk requiring further investigation. Reviewers need to understand both.
Acceptance Has Boundaries
There is no single finding that makes a NAM suitable for every product or use. Evidence supporting its use for skin irritation does not establish that it can assess reproductive toxicity or interactions between organs.
FDA’s new database of NAM use examples makes this explicit. Inclusion illustrates use in a particular application. It does not approve the method for other uses or promise acceptance in a future submission.
Formal validation is also distinct from consideration of evidence in an individual development program. The draft guidance allows a fit-for-purpose NAM that has not been formally validated to address a specific concern within the overall weight of evidence. The method’s contribution must be assessed alongside the other information about the product. The guidance also bounds itself: it addresses NAM data submitted in support of drug applications to CDER, not the use of NAMs in drug discovery or the separate qualification of drug development tools.
There are questions for which animal studies remain important. The rule identifies complex physiological interactions involving hormones, neurotransmitters, cytokines, and communication between organ systems. A method that represents only selected parts of that biology can leave important effects unexamined.
Recognizing those limits is part of deciding what additional testing is needed.
Human Testing Still Has Its Own Purpose
The rule is not confined to the initial investigational new drug application. It also changes terminology in provisions covering marketing applications and safety reporting. Its scope excludes the Animal Rule provisions and the orphan-drug regulations.
But changing how nonclinical safety evidence is generated does not remove the need for the clinical evidence required for approval.
Before initial human exposure, researchers use nonclinical evidence to assess whether the proposed trial can proceed with an acceptable basis for safety. Clinical investigations then provide evidence from people. Further nonclinical work can support later decisions as development continues.
A favorable model result is therefore one contribution to a continuing evaluation. Its meaning depends on the proposed exposure, the uncertainty that remains, and the other evidence available.
