FDA wants you to 'show your work'
Regulators are pushing teams to validate new lab and modelling methods with clear evidence linking how a method is used to its performance, not just flashy claims. A recent primer on the FDA’s draft guidance for validating new approach methodologies argues the agency expects context-of-use, reproducibility and traceable performance for methods to inform decisions (news-medical.net). That means digital twins, ML classifiers or automated assays that can’t document training data, boundaries and drift will struggle in regulatory review, making provenance and lifecycle control operational priorities (news-medical.net).
For decades, drug companies could hand regulators a stack of animal studies and move on. On March 18, 2026, the Food and Drug Administration said that if you want to use newer tools instead, you now need to prove exactly what the tool is for and how well it performs. (fda.gov) Those newer tools are called new approach methodologies, which is a long label for things like lab-grown tissue models, computer simulations, and cell-based tests meant to predict how a drug will behave in people. The agency’s draft says these methods can support drug development, but only if they are validated for the decision they are supposed to inform. (federalregister.gov) The basic problem is that animal studies are slow, expensive, and often a bad stand-in for humans. The Food and Drug Administration says animal tests can take years, cost a lot because animals must be housed and cared for, and often do not reveal the biological mechanism behind the harm they detect. (fda.gov) That is why regulators have been moving toward more human-based systems. In an April 2025 roadmap, the agency said it wanted a stepwise shift toward organ-on-a-chip systems, computational modeling, and advanced laboratory assays that could reduce animal testing in preclinical safety work. (fda.gov) The new draft guidance is the rulebook for that shift. The Food and Drug Administration says it sets four core validation principles: context of use, human biological relevance, technical characterization, and fit for purpose. (fda.gov) Context of use is the agency’s way of asking one plain question: what exact job is this method doing. A model built to flag liver toxicity early in preclinical screening is not automatically good enough to replace a broader safety package in a regulatory filing. (fda.gov) Technical characterization is the “show your work” part. The Federal Register notice says the guidance focuses on study design and reporting principles, which means companies need records that let reviewers trace how a method was built, tested, and bounded before trusting its output. (federalregister.gov) That lands especially hard on software-heavy tools. If a machine learning classifier cannot document its training data, if a digital twin cannot define where its predictions stop being reliable, or if an automated assay cannot show the same result across runs, the validation file starts to look thin. (news-medical.net) The agency is not banning flashy new models. It is telling companies that a sophisticated model without reproducible evidence is like a calculator that gives the right answer sometimes but cannot show the equation. (fda.gov) This is still a draft, and the public comment window runs until May 18, 2026. The Food and Drug Administration is asking industry to help shape the final version, but it has already made the direction clear: newer methods can move drugs forward only if sponsors can tie performance to purpose, in writing, for regulators to inspect. (federalregister.gov)
