Regulators Build an AI Arsenal: Unpacking the EMA 2025 Observatory Report

Sam: The single most interesting finding here is that the regulators are not just reading your data anymore. They are actively deploying their own suite of artificial intelligence tools, identifying 61 specific internal use cases to interrogate pharmaceutical submissions.

Maya: It completely shifts the balance of power in drug approvals. Today we are doing a deep dive into the 2025 AI Observatory report, which was published by the European Medicines Agency.

Sam: And before we get into what this means for your clinical practice or your boardroom strategy, a quick reminder that our voices are AI-generated and this is a summary of a public document.

Maya: This report is produced annually by the Network Data Steering Group. Their mandate is to monitor information on activities, trends, and emerging AI domains across the European Medicines Regulatory Network. They are looking at both human and veterinary medicines.

Sam: The forewords by Karl Broich and Peter Arlett set a very distinct tone. They say that in 2025, AI was moving from concept to practice. Industry is integrating it, and regulators are adopting it to enhance insight and efficiency.

Maya: Right, the foundational rules are now in play. The document notes that the EU AI Act actually entered into force in 2024, but 2025 is the year it moved into practical application. That started with prohibited AI uses and AI literacy obligations.

Sam: And then it moved into general-purpose AI obligations and governance structures. So for pharma companies and medtech developers, the theoretical phase is entirely over. You have to show how your tools are transparent, fair, and aligned with public trust.

Maya: Exactly. And the primary way the agency is building its knowledge base is through early interactions with applicants. The report repeatedly highlights how new applications of AI are primarily discussed during early regulator-stakeholder interactions.

Sam: They mention very specific forums for this. The Portfolio and Technology Meetings, which they abbreviate as PTM, and the Innovation Task Force meetings. Why are developers bringing their AI to these groups before filing a formal submission?

Maya: Because they need to gauge the likely regulatory acceptability of these tools early on. If you build a drug development pipeline around an AI model that the regulator fundamentally distrusts, your entire investment is at risk. These early meetings give the network visibility into the emerging pipeline.

Sam: Let us walk through the lifecycle, starting with pre-clinical development. How is AI changing the way we look at drug candidate identification and safety before a drug ever reaches a human trial?

Maya: The report highlights a shift toward coupling New Approach Methodologies with machine-learning algorithms. These include things like organ-on-chip models, high-content imaging, and systems biology. The goal is to interpret complex biological responses better than traditional methods.

Sam: And the ultimate aim is reducing the reliance on animal studies while actually improving human relevance. They also mention that in the domain of genotoxicity assessment, in silico models are formally embedded in regulatory frameworks like ICH M7.

Maya: Yes, they specifically point to machine-learning-based QSAR models. These are used alongside expert rule-based systems to predict the mutagenic potential of impurities. Those rapid, data-driven assessments directly inform the weight-of-evidence evaluations that regulatory assessors perform.

Sam: Moving into clinical development and clinical trials, the scale of what is being attempted is massive. Let us start with just getting a trial off the ground. The report states AI is being used to support clinical trial country and site selection.

Maya: It goes further than just picking a site. Machine learning is being used to predict enrolment rates at potential clinical trial sites. They want to identify the sites most likely to complete a study faster through faster enrolment of patients.

Sam: That is a pure operational efficiency play, but it saves millions of dollars. They also talk about using natural language processing to extract external and internal knowledge to develop better protocols in less time.

Maya: And once the trial is running, the way we measure outcomes is being transformed. The report has a huge section on medical imaging. AI is automating endoscopic scores and automating arthritis scoring.

Sam: Wait, really? They are letting AI grade the severity of diseases from images? What specific diseases are they looking at in these early meetings?

Maya: A wide variety. The Annex lists machine learning algorithms for scoring radiographic progression in Psoriatic Arthritis, and algorithms for scoring endoscopy videos in Ulcerative Colitis. They even mention deep learning based algorithms to analyze kidney segmentations.

Sam: And this is not just theoretical anymore. The report explicitly mentions AIM-NASH, which is an AI-based measurement of NASH in liver disease to determine disease activity. It says a qualification opinion was published in April 2025.

Maya: That is a critical milestone. A published qualification opinion from the agency means that the methodology is officially recognized as acceptable for a specific intended use. It is a massive green light for the developers of that tool and sets a precedent for the industry.

Sam: The report also delves into clinical outcome prediction. There are tools called MesoNet and HCCnet that perform statistical adjustment on deep learning prognosis covariates obtained from histological slides.

Maya: The underlying goal of those models is to reduce outcome variability. If you can produce baseline prognostic covariates to use in treatment effect analyses, you get a much clearer picture of whether your drug is actually working. They are also doing this for Major Depression Disorder.

Sam: They mention building prognostic values of response for multiple diseases from clinical trial data. The agency notes this can explain the variability of clinical endpoints and hence improve clinical trial sample size planning.

Maya: Which is exactly what boardroom leaders want to hear. Smaller, more precise trials. They are also exploring something very interesting for trial participants directly, which are generative AI-based assistants.

Sam: I saw that. They call them digital clinical assistants. It is new software technology enabling clinical trial participants to complete various clinical tasks through interaction with generative artificial intelligence.

Maya: And we cannot forget External Control Arms or Synthetic Control Arms. The report notes that developers are using external and historical patient data to establish comparator arms. They want to avoid recruiting new patients to take a placebo if they can just use historical data.

Sam: But that relies entirely on the quality of the data and the robustness of the model. That transitions perfectly into the manufacturing domain, where the models get even more complex. The report focuses heavily on pharmaceutical process models and digital twins.

Maya: The Quality Innovation Group, or QIG, had several meetings with applicants on these process models. They are looking at AI applied to CAR-T cell manufacturing and AI platform approaches to characterize cell-derived products.

Sam: They even mention a computer vision AI algorithm for the automatic counting of microcarriers and determining colonization percentage. So instead of a human peering through a microscope on the factory floor, a camera and an algorithm are doing it.

Maya: Yes, and they are using AI to optimize manufacturing processes by predicting maintenance needs and improving quality control. This is supposed to ensure consistent quality and reduce waste.

Sam: One of the most audacious applications discussed is Predictive Stability Modelling and Shelf-life Testing. Typically, you have to put a product in a room and wait to see how long it takes to degrade.

Maya: But now, machine learning is used to accurately predict attribute values using various product characteristics and initial lot-release data values, entirely without long-term stability data. They want to prove shelf-life mathematically from day one.

Sam: That is incredible, but what if the AI is wrong? If a stability prediction fails, people could receive degraded medicine. How does the agency view the risk of using AI in manufacturing?

Maya: The report is actually quite pragmatic about that. It notes that the guardrails inherent in Good Manufacturing Practice, or GMP, provide risk mitigation for many of the general risks of AI. Because the environment is so tightly controlled, the variables are limited.

Sam: They do mention traceable robotics to be used in aseptic environments to reduce human intervention, using deep learning for sensing and object recognition. Less human intervention means less contamination risk.

Maya: Quick note before we carry on. This spot is open for a sponsor. If your company builds or sells AI for healthcare and wants to reach the clinicians, health-system leaders and industry teams who listen to this show, the link to our sponsorship page is in the show notes.

Sam: And now, back to the document.

Maya: Exactly. Now, let us look at post-marketing authorization. Once the drug is approved and patients are taking it, how does the agency monitor safety? The report details AI approaches being explored in the generation of real-world evidence.

Sam: They mention using AI to streamline data cleaning and create synthetic control arms. But they also highlight social media screening for pharmacovigilance. They are monitoring social media accounts to identify potential safety cases.

Maya: It is a necessity. Social media screening presents unique challenges, primarily the massive volume of unstructured data. Companies have developed AI approaches that yield faster detection of adverse events and enhanced patient safety compared to traditional approaches.

Sam: And when those adverse events are detected, they have to be processed. AI is being explored to support the intake and management of individual case safety reports. That includes information extraction, translation, narrative generation, and medical review.

Maya: But here is the turning point of the entire document. We have spent all this time talking about what the developers are building. The report makes it clear that regulators at the national and network levels are building their own tools to keep up.

Sam: This is the most fascinating part. The Network Data Steering Group organized workshops with national competent authorities and collected 61 use cases across four main AI areas. The regulators are staffing up with AI.

Maya: Those four areas are drafting and summarization, validation and quality assurance, knowledge mining, and other diverse use cases. They are literally building an ecosystem of AI assistants to optimize knowledge retrieval.

Sam: And to ensure they do it well, they agreed to establish an EMRN Prompt Community in 2025, starting as a pilot. They are treating prompting as a core regulatory competence. They want a collection of validated and reusable prompts.

Maya: They have already rolled out tools. Scientific Explorer has been available since March 2024. Initially, it improved searches related to scientific advice procedures. But in 2025, its functionalities were expanded to include information on initial marketing authorisation applications of human medicines.

Sam: Then there is the AI@MPA toolbox, which is provided by the Swedish Medical Products Agency. It is a web-based suite for EU regulators. In 2024 it had 6 applications, helping assessors navigate product information and identify guidelines.

Maya: But in 2025, they expanded it significantly. They added a generative AI tool called REGULUS for document processing and answering regulatory questions. They added PACKSIM for drug-package similarity search. And most impressively, they added SCHEMA.

Sam: SCHEMA stands for SMILES chemical embeddings map. What does that actually mean for an assessor looking at a new drug application?

Maya: It means they are not just relying on the safety narrative the applicant provides. SCHEMA enables read-across of preclinical and clinical safety data to new small molecules. The assessor can mathematically map the chemical structure and automatically pull historical safety data from similar molecules across their database.

Sam: That is a totally different level of independent verification. The agency is using AI to fact-check the applicant's AI. And to manage all of this, the NDSG adopted the Network AI Tools framework and catalogue to share these tools across the whole network.

Maya: This transformation requires massive collaboration. The report details intra-EU cohesion efforts and international convergence. The EMA and the U.S. Food and Drug Administration jointly identified ten principles for good AI practice in the medicines lifecycle.

Sam: Those principles guide the use of AI from early research right through to manufacturing and safety monitoring. It is clear they want global strategic alignment. The report notes that 4 ICMRA Regulatory Forums were organized in 2025.

Maya: During those forums, the ICMRA AI Steering Committee exchanged information on things like using AI chatbots as the first line interaction for clinical trial applications, and approaches for assessing risk in machine learning based tools.

Sam: The EMA also chaired the EU Agencies Network Working Group on AI. Their annual plenary meeting took place in June at the EMA premises. It was massive, with participation of EU Agencies and Joint Undertakings sharing expertise and showcasing use cases.

Maya: They discussed AI systems procurement, the EUAN staff exchange programme, and assessed needs related to the EU AI Act implementation. They even discussed the GPT@EC pilot plans, looking deeply at data protection, AI Act compliance, and intellectual property considerations.

Sam: And the agency is not just collaborating with other regulators; they are funding massive scientific research projects. Annex 3 is a comprehensive list of EU-funded initiatives exploring the application of AI across the lifecycle.

Maya: The breadth of these initiatives shows where the agency sees the future of medicine. In the pre-clinical domain, there is a project called QUANTUM-TOX applying AI to develop computational toxicology, and another called Ai4Cilia analyzing ciliary beat defects for drug discovery.

Sam: They mention TClock4AD, which is developing artificial intelligence strategies for new circadian clock drug candidates targeting Alzheimer's Disease. And NEWROAD, an open in silico platform for repurposing drugs in rare and paediatric cancer research.

Maya: In clinical development, the Horizon 2020 Framework Programme funded BRAINTEASER. That project integrates large clinical datasets with novel personal and environmental data collected using low-cost sensors and apps from patients with amyotrophic lateral sclerosis and multiple sclerosis.

Sam: There is also OPTIMA, aiming to be the first interoperable and GDPR compliant European real-world oncology data platform. And KATY, an AI-empowered Personalized Medicine system bringing medical knowledge directly to clinicians.

Maya: I was particularly struck by Histotype Px. They are applying AI to digital pathology for the stratification of colorectal cancer patients into low, intermediate, or high risk prior to chemotherapy. And AI4LUNGS is building models to improve patient stratification for respiratory diseases.

Sam: In manufacturing, they highlight AIDPATH, which applies AI to the decentralized production of advanced therapies in hospitals. And AiPSC, an AI-powered platform for autologous iPSC manufacturing. The momentum is undeniable.

Maya: But the report is very explicit about what is still missing. Despite the breadth of scientific research, they say gaps remain in the area of AI tools for regulatory assessment. The science is advancing, but the regulatory-grade acceptability is lagging.

Sam: What exactly do they mean by regulatory assessment gaps? What do they need that they do not have?

Maya: They specifically cite the need for model validation, explainability evaluation, and comprehensive audit frameworks. They also noted a lack of veterinary AI initiatives, and the need for continuous AI monitoring. Effective translation of findings into regulatory practice is critical.

Sam: To guide researchers and funding bodies to fill those exact gaps, the Network Data Steering Group adopted what they call Network AI research priorities. They identified 7 distinct domains that need urgent attention.

Maya: Those 7 domains form a roadmap for anyone building in this space. They cover research integrity and intellectual property, meaning how to manage proprietary rights for models and data. They cover the accuracy and reliability of tools, including explainable outputs and robustness.

Sam: They also focus on data governance, confidentiality, and consent. Secure, lawful, and ethical handling of patient data is paramount. And regulation and oversight, looking at accountability and liability frameworks.

Maya: The final three domains are ethics, fairness and bias prevention to avoid discrimination; resources and support for AI use; and the impact on jobs and skills. They are actively studying how AI reshapes roles and demands new workforce transitions.

Sam: The agency even launched the European Platform for Regulatory Science Research in March 2025 to bring academia and regulators together. They want to proactively inform funding bodies on regulatory needs rather than waiting to see what gets built.

Maya: To give a crisp recap: the EMA is shifting from just reading AI guidelines to actively deploying 61 internal AI use cases, building tools like SCHEMA to fact-check safety data, and collaborating globally to set the guardrails.

Sam: Exactly. The balance of power in drug approvals is fundamentally changing, and developers must be ready for highly sophisticated regulatory scrutiny. Be sure to read the full 2025 AI Observatory report, which is linked right to the source document in the show notes.

Maya: And as a final reminder, this podcast is for informational purposes only and is not medical advice. See you next time!