It’s Not About Replacing Mice With Algorithms: FDA Sets Validation Thresholds for AI Biosimulation

Alternatives to animal testing are shifting from a laboratory vision to an issue that drug regulators must directly address. When algorithms can simulate organ responses and organoids can reproduce certain characteristics of human tissues, the next hurdle is no longer simply whether these approaches are feasible, but whether their results are sufficient to support a new drug’s entry into human trials and even become part of regulatory decision-making.

*Nature Biotechnology* reviewed three technological paths that are gradually converging. Digital twins seek to simulate how drugs act in a specific organ or the human body using physiological data and mechanistic models. Multi-agent virtual scientists can divide tasks such as proposing hypotheses, designing analyses, and checking inferences. AI-enhanced organoids combine human-derived three-dimensional tissues with imaging and molecular data analysis to identify efficacy or toxicity phenotypes. These technologies could help developers screen out high-risk drug candidates, investigate mechanisms of action, and select doses and trial designs that are more promising for clinical development.

However, an impressive simulation does not equal credible biological evidence. Digital twins may be limited by the representativeness of training data, model assumptions, and errors in extrapolation across scales. Organoids generally cannot fully reconstruct immune, metabolic, and multi-organ interactions. Even when multi-agent systems can produce seemingly coherent research plans, they may collectively amplify flawed premises. If these tools are to replace rather than merely supplement animal data, they must demonstrate that their results are reproducible and can predict human responses relevant to a specific regulatory question.

A new methodology draft released by the FDA in March is making these requirements more concrete. The draft covers organoids, organs-on-chips, computer simulations, and other in vitro or computational methods, and sets out four validation principles: first define the context of use; demonstrate relevance to human biology; complete assessments of technical characteristics and reproducibility; and finally confirm that the tool is fit to support the intended regulatory decision. In other words, a model that has worked for one disease or organ will not automatically apply to another indication, toxicity endpoint, or drug type.

The regulatory infrastructure is also beginning to take shape. In its annual progress report in April, the FDA said it had expanded weight-of-evidence assessments combining in vitro testing and computational toxicology, established cross-center scientific review and qualification pathways for innovative tools, and qualified the first AI-based drug development tool. However, the qualification of a single tool does not mean that all AI models will be accepted. The FDA still recommends that developers consult the relevant review division early, because the level of acceptance depends on the disease, organ, endpoint, and actual use.

Early FDA research has already demonstrated both the potential and the limitations of this approach. AnimalGAN uses existing animal testing data to generate synthetic data for untested chemicals, while SafetAI uses deep learning to predict toxicology endpoints, including drug-induced liver injury. Such systems can repurpose vast amounts of historical data, but they may also inherit the biases and gaps in older datasets. The turning point for AI biosimulation therefore lies not in declaring that animal testing is about to disappear, but in building, item by item, traceable, comparable, and fit-for-purpose validation evidence—turning the claim of being “closer to humans” from a technological assertion into a reviewable scientific fact.