The Medicines and Healthcare products Regulatory Agency has opened a call for evidence examining how artificial intelligence is being used to assess medicines safety, including barriers around model validation, data access and regulation.
The exercise forms part of the regulator’s Beyond ADMET programme and will inform development of a regulatory sandbox focused on AI-enabled medicines safety. Organisations have until 22 December to respond.
The MHRA is seeking evidence from companies and researchers involved in toxicology, preclinical development, clinical research, biotechnology, life sciences and artificial intelligence. Organisations do not need to be using the technology already, with the regulator also interested in future plans and unmet needs.
Potential applications include prediction of absorption, distribution, metabolism, excretion and toxicity — collectively known as ADMET — alongside analysis of clinical and real-world evidence. The objective is to understand whether AI can identify safety risks earlier and improve how medicines are assessed across different patient groups.
Medicines development is an attractive target for AI because conventional discovery and testing can be expensive and slow. A system able to eliminate weak candidates earlier could reduce unnecessary laboratory and clinical work and allow researchers to concentrate resources on compounds with stronger prospects.
The scientific challenge is considerably harder than many commercial AI applications. Biological systems are complex, and a model can identify relationships in historical data without proving that a candidate will behave safely in humans.
Regulators therefore need evidence about reproducibility, uncertainty and external validation rather than simply headline accuracy on a development dataset. Models also have to perform across populations that may be poorly represented in the information on which they were trained.
That creates a different oversight problem from conventional software. An AI system can change when its underlying model or dataset is updated, meaning performance at the time of initial assessment may not describe later versions.
Developers consequently need governance around model changes, while regulators need to determine when an update is sufficiently material to require additional validation. Similar questions are already affecting other areas of healthcare AI.
Data access presents another constraint. Pharmaceutical companies, universities, health systems and regulators hold large volumes of safety information, but the data can be fragmented across incompatible systems and subject to privacy, commercial and legal restrictions.
Larger and more representative datasets could improve model performance, yet combining information across organisations introduces questions over ownership, consent, confidentiality and cybersecurity. Technical capability alone does not resolve those governance issues.
The proposed sandbox is intended to provide a controlled environment in which developers and regulators can test approaches before requirements are fully established. Regulators have increasingly used sandboxes where emerging technologies develop more quickly than conventional rulemaking and practical evidence is needed to shape future oversight.
The MHRA exercise sits within a wider reassessment of AI regulation in healthcare, where policymakers are trying to encourage useful innovation without lowering evidential standards for technologies that can influence patient safety.
The call for evidence introduces no immediate new compliance obligation. Its significance lies in the rules that may follow. Businesses developing AI for drug safety have an opportunity to influence how validation, data access and model governance are treated before those systems become more widely embedded in medicines development.
You must be logged in to post a comment.