MHRA seeks evidence on AI tools that could reshape medicines safety assessment
The Medicines and Healthcare products Regulatory Agency has opened a call for evidence on artificial intelligence in medicines safety, framing the exercise as part of a broader effort to move beyond traditional ADMET assessment. The initiative asks organisations to describe how AI is being used, developed, or planned across the medicines pipeline, with particular attention to model development, validation, access to data, data sharing, and the regulatory questions that arise when algorithms are folded into safety evaluation.
The invitation lands at a moment when preclinical screening and translational science are both under pressure. Conventional approaches to absorption, distribution, metabolism, excretion and toxicity have long been useful, but they remain constrained by the limits of experimental models and the variability of human biology. A compound that looks acceptable in a controlled preclinical setting can behave very differently once it encounters comorbidity, polypharmacy, organ impairment, age-related physiology, or the genetic diversity seen in routine care. The regulatory logic behind the survey is clear: if machine learning and related AI methods can integrate more dimensions of biological and clinical data, they may identify safety concerns earlier and more accurately than older workflows alone.
The deadline for responses is 22 December 2026. That date gives industry, academic groups, and regulators a defined window to describe what is already working and what remains unresolved.
Why the call for evidence matters for translational science
The regulatory framing goes beyond simple automation. The survey focuses on whether AI can combine traditional nonclinical evidence with clinical and real-world data to predict how a medicine will behave across different populations and disease states. In practical terms, this is the central challenge of modern drug safety science: a model that performs well in a narrow dataset may fail when exposed to patients with renal dysfunction, hepatic impairment, inflammatory disease, or overlapping therapies.
That problem is not abstract. Safety signals often emerge only after a medicine is used in broader clinical settings, where adherence varies, background treatment is messy, and patient characteristics are far less uniform than in early development. AI systems promise to identify patterns that may be too diffuse for conventional analysis, but those systems also inherit the weaknesses of the data that train them. If datasets are incomplete, inconsistent, biased, or poorly linked, the output can appear sophisticated while remaining clinically fragile.
The MHRA and the Regulatory Innovation Office are asking respondents to address those exact tensions. The consultation seeks practical examples from toxicology, preclinical research, clinical testing, life sciences, biotechnology, and adjacent AI sectors. It also invites views from organisations that are not yet using AI for medicines safety. That broader scope is important because regulatory readiness depends not only on current users, but also on the groups that can identify barriers before they become embedded in practice.
Data quality and validation sit at the center of the discussion
Validation is the pressure point that will determine whether these tools have regulatory value. AI can only improve medicines safety if its predictions can be checked against outcomes that matter in the real world, including adverse events, organ-specific toxicity, and unexpected behavior in special populations. Without rigorous validation, a model may generate plausible predictions that are difficult to defend in submissions or regulatory review.
The survey also points to a second bottleneck: access to data. Medicines safety models often depend on diverse datasets that span preclinical findings, trial records, and real-world evidence. Those sources are rarely assembled in a single, tidy system. Data sharing introduces technical, legal, and governance challenges, especially when information must move across institutions or national boundaries. For regulators, those barriers are not peripheral. They are often the difference between a concept that can be tested and one that remains confined to pilot work.
Model development itself raises additional questions. Different AI methods may be suited to different tasks, from toxicity prediction to signal detection and exposure modeling. Yet the regulatory burden does not disappear when the model becomes more complex. In fact, interpretability may become even more important when decisions affect which candidates advance, which warnings are investigated, or which populations need extra monitoring.
A prediction that cannot be explained is hard to regulate and harder to trust.
A sandbox approach is meant to inform future regulation
The findings from this call for evidence will feed into the Beyond ADMET: AI for medicines safety regulatory sandbox, announced earlier in 2026. That sequencing is notable. Rather than writing rules in the abstract, the MHRA and RIO are gathering operational detail first: who is using AI, where the technology is already embedded, what types of validation are considered acceptable, and which technical or governance hurdles most often block deployment.
A sandbox model is especially relevant in medicines safety because the consequences of weak oversight can be severe. If AI is used to prioritize toxicology experiments, support early decision-making, or flag safety concerns from mixed datasets, regulators need to know where human review remains essential and where automation may be acceptable. The consultative process may also shape workshops and later engagement with innovators and researchers, giving the agency a route to test assumptions in a more interactive setting.
For organisations working in this space, the request is straightforward but consequential. The MHRA wants examples of current practice, unmet need, and future plans. It wants to know where AI is already helping, where it is still aspirational, and which obstacles keep promising methods from entering routine use. That includes not only technical issues, but also the regulatory considerations that determine whether a model can be used with confidence in a safety-critical setting.
The unanswered questions are operational, not philosophical
The consultation does not ask whether AI should play a role in medicines safety. It assumes that role is coming and asks how best to govern it. The practical questions are sharper: Which datasets are robust enough for model training? How should external validation be performed across populations and clinical conditions? What level of evidence is needed before a predictive system influences drug development decisions? Which aspects of safety assessment still require human judgment because the data are too sparse, too noisy, or too context-dependent for automation?
Those issues will matter long after the survey closes. The regulatory path for AI in medicines safety will depend on whether the field can generate reproducible evidence, share it responsibly, and keep pace with changing clinical reality. For now, the MHRA has opened a formal channel for that evidence to reach the people shaping future guidance. The closing date is fixed, but the larger question is whether the sector can produce the kind of validated, regulator-ready science that turns AI from a promising concept into a dependable safety tool.