Let me start with the obvious. AI has dominated the conversation for the last three years, and as the editor that creates a real dilemma. When everything is labeled AI, an AI issue can feel predictable before you turn the first page. I did not want that. I wanted this issue to be useful and grounded, to give you practical takeaways you can apply with your team, and to spotlight a few unusual applications that challenge how we think about the work.

So we made a clear choice to focus on practice, not slogans. You will see how pairing subject-matter experts with AI changes outcomes in manufacturing and quality; when experts frame the problem and set the standards, AI stops being a toy and becomes a tool that shortens reviews, surfaces specific gaps, and turns weeks of document preparation into days. Beyond Generic AI: How Domain Expertise Creates Breakthrough Tools for Pharmaceutical Operations. We also looked at supply chains, where startups run on cash and time, and better forecasting can prevent shortages, coordinate vendors, and protect both. Operational Intelligence: How AI is Rewriting the Playbook for Supply Chains in MedTech and Biotech Startups.

On the clinical side, we move past the usual headlines and get into design choices. One feature examines digital twins that can right-size control arms without compromising statistical power, which is a practical lever on timelines and budgets. Digital Twins in Clinical Trials. Another looks at the EU AI Act and what it means when your trial relies on systems that regulators consider high risk; the takeaway is straightforward. If AI touches regulated work, build for transparency, documentation, and human oversight you can defend in an audit. EU AI Act and Clinical Trials. To balance optimism with discipline, we include a perspective on where AI helps regulatory planning today and where it does not, so you can set credible expectations with your stakeholders The Future of AI in Regulatory Planning: Progress with Caution.

I also wanted examples that do not look like yesterday’s slide deck. One story covers early cancer detection that blends canine olfaction with a decision platform; it is unconventional, which is exactly why the operational and validation model is worth your time. Harnessing Nature and Nurture: How SpotitEarly’s Bio-AI Hybrid Platform Redefines Early Cancer Detection. Another centers on patient communication, arguing that the first hour after diagnosis sets the tone for months and that clinically validated, plain-language guidance should be treated as part of care, not an add-on. Clear, Clinically Validated Communication: Transforming Patient Care.

To help you operationalize all this, we include a compact toolkit for AI governance and traceability so you can measure, explain, and reproduce model behavior in environments that regulators actually understand ==[link: Toolkit: TruLens, Weights & Biases, ClearML]==. And we step outside software with a piece on protein design that starts from nature’s own blueprint and moves toward formulations that hold up in the real world; different field, same principle of starting from what works and building carefully ==[link: Human Milk as Nature’s Gold Standard]==.

My goal with this issue is simple. Less noise and more signal. If you are building, testing, or scaling in biotech or medtech, I hope these stories help you move faster and safer. Our role at ADRES.bio is to make that practical; we help teams turn AI into inspection-ready operations with people who know the science and the systems. If something here sparks a question, tell us. And if you have a case study for the next issue, send it our way.

Exploring Challenges and Approaches for Ethical and Effective AI in Healthcare

The rapid growth of health data from electronic health records, genomics, and wearable devices has created unprecedented opportunities in medicine. Artificial Intelligence is increasingly viewed not merely as a support tool but as a cornerstone for future diagnostics and personalized treatments. This shift, while promising, introduces significant challenges: ensuring transparency, preventing bias, and integrating technology with human clinical judgment.

One promising approach gaining traction is phenotype-driven medicine. Instead of grouping patients by broad disease labels, this method identifies smaller, more meaningful clusters based on biological markers, behaviors, and health histories. Such clustering can guide highly personalized treatments and improve outcomes compared to traditional, one-size-fits-all models.

A practical illustration of this can be found in initiatives like Obesio, which applied phenotype clustering to weight management. By identifying six distinct metabolic profiles, Obesio enabled tailored interventions that combined medication, nutrition, and exercise specific to each profile. Results from early pilots showed over 30% greater weight loss and improved patient adherence compared to conventional methods, highlighting the potential impact of this approach.

Introducing AI tools into clinical workflows requires more than technological sophistication; it demands careful consideration of ethics and operational realities.

A useful framework to guide this integration is the Human-Machine-Doctor Triangle, which focuses on three critical dimensions:

1. Explainability (XAI): Physicians must understand why an AI makes a recommendation. Transparent models help build trust and support informed clinical decisions.
2. Bias and Equity: AI models must be trained and audited with diverse data to avoid reinforcing existing health disparities. Ongoing evaluation across populations is essential.
3. Data Privacy and Patient Autonomy: Trust is earned through transparency and compliance with global standards like GDPR and HIPAA, ensuring patients know how their data is used and protected.

These principles aim to ensure AI enhances rather than replaces the human elements of care-empathy, judgment, and the patient relationship. When implemented thoughtfully, AI can surface insights from data too complex for humans to process, allowing clinicians to focus on what they do best: caring for patients.

As healthcare moves toward proactive, predictive, and participatory models, frameworks like phenotype-driven medicine and the Human-Machine-Doctor Triangle offer practical pathways forward. They demonstrate that the future of personalized medicine will rely not just on technology, but on the careful blending of computational power, ethical oversight, and clinical wisdom.

Let’s do something a properly trained writing AI probably would not do: Start with an I statement.

I get the feeling that each time ADRES reaches out to me about possibly contributing an article to the BioStartup Standard, it turns into me writing about parts of my personal journey in the biotech space.

It is the same again, this time, because four (widely spaced) personal events, or perhaps rather encounters, over the course of that journey, together inform this piece, which, stripped of those anecdotes, boils down to little more than a small piece of advice regarding training: Training not, as the context of AI may suggest, in the sense of feeding training data into an AI model, but training in its traditional sense – the training of people.

1. The first touchpoint was reading an article about knowledge loss (organizational forgetting) in the chemical industry. This was more than 15 years ago, and I did not keep a copy of the article, so unfortunately, I cannot attribute the exact source. Part of the methodology included interviews with retired lead engineers from several chemical plants and the role they continued to play as consultants post-retirement. It highlighted how there was no impact on routine operations with the loss of key knowledge assets (the lead engineers), but that that changed as soon as troubleshooting was required, be it due to quality issues, breakdowns, or changes such as planned expansion or process improvement.

• Other than the first, the second touchpoint was directly related to biotech and to AI – or rather to Natural Language Processing (NLP), i.e., one of the key concepts in machine learning and language models, because back then – in the early 2010s – no one I knew called it AI yet. But the ideas were already there and I was discussing their potential application to biotech (specifically, to the analysis and presentation of data from clinical trials) with friends in Cambridge who were doing NLP research. While we could already envision, if dimly, what would be possible in the future (was is possible now!), in the short- to midterm we saw that the limitations of the (then available) technology would place it firmly as a tool for a human expert, like an advanced word processor or statistical programming suite.

• Fast forward to 2025, and we have AI established and growing in importance across industries (including, of course, biotech). And along with that we have a growing body of criticism as well, which is where the third touchpoint comes in: A couple of weeks ago (end of June 2025, that is), a friend recommended a draft paper[1] to me, covering a study on neural and behavioural consequences of using AI assistance in (academic) writing tasks. The authors’ concluded that aside from positive effects the use of AI also “came at a cognitive cost”, impacting critical evaluation of AI outputs and potentially reinforcing “echo chamber” environments in which outputs from AI systems get critically checked less and less as their users get primed by previous exposure.

• Then, shortly thereafter, the final piece to this puzzle, the one that made everything click into place, came into play when colleagues at ADRES reached out with the call for contributions to the issue of the BioStartup Standard you are currently reading. And right there, in the middle of the technical guidelines for submitting an article, I read “AI tools can assist, but substantial revision and personalization are required” and found that mildly funny – the call for contributions to “the AI issue” was critical of relying fully on AI. Initially, my somewhat vague intention had been to write about implementation of “behind the firewall” systems in small scale organisations or something similar more operations oriented. But I felt myself constantly drawn back to this critique of AI in a call for AI and it got me thinking in an entirely different direction. One by one the above memories came up: My first – abortive (I would be lying if I said that we implemented anything of what we discussed in Cambridge) – concepts for utilising AI in trial analysis and reporting as a tool for human experts; the recent paper on cognitive cost of using AI – specifically in an educational (learning!) setting; the long ago read about organizational forgetting caused by personnel turnover; it all started to fit together.

Let me pause here briefly to state (if that did not become clear from the Cambridge anecdote) that I am not an AI-luddite who tries to warn you about how dangerous this technology is and that you better not use it. We are using it. And we should be using it. It is a powerful tool, as I am sure a lot of the colleagues contributing to this issue will highlight in their own articles.

As we are adding new and powerful tools to your toolbox, we need to make sure to also have the right users for these tools, not just the right tools themselves, and that also means not neglecting the training of your next generation of users – not just in using the tools, but in the fundamentals.

The current generation of professionals in our space has still acquired their skills and experience outside an AI echochamber, they are experts able to deliver without AI support, who become further empowered by new AI tools, and are able to critically review what a system delivers, feeding into a continuous improvement cycle.

But this generation is not here to stay forever. What is needed, thus, is to ensure that the next generation, as well, will understand the underlying science and processes, and often enough the art and craftsmanship to do the same – to function and deliver without AI, to make the most use out of the AI systems available, to check whether the systems are performing, and to improve them going forward.

Invest in AI. But do not neglect to invest in people.

AND not OR.

1 N Kosmyna, E Hauptmann, YT Yuan, J Situ, X-H Liao, AV Beresnitzky, I Braunstein, P Maes, ‘Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task’, https://arxiv.org/abs/2506.08872 (retrieved 30-Jul-2025)

The frontier of medical diagnostics is rapidly evolving, driven by the convergence of diverse scientific disciplines. In this exciting landscape, where biology meets artificial intelligence, a revolutionary approach to early cancer detection is emerging. SpotitEarly is pioneering a bio-AI hybrid platform that integrates the unparalleled olfactory capabilities of trained detection canines with advanced Deep Learning algorithms, transforming how we identify cancer at its earliest, most treatable stages. This innovative synergy promises to transcend traditional diagnostic limitations, offering a scalable, non-invasive, and highly accurate solution for multiple cancer types screening.

The Unrivaled Nose: Nature’s Diagnostic Powerhouse

At the heart of SpotitEarly’s technology lies the extraordinary biological ability of detection animals, particularly canines, to discern minute concentrations of Volatile Organic Compounds (VOCs). Cancerous physiological processes produce detectable odorants, or VOCs, which are exhaled in breath, sweat, saliva, urine, and other bodily fluids. Canines possess extremely sensitive olfactory receptors, allowing them to pick out specific scent molecules, even at low concentrations. Unlike conventional diagnostic tools that may struggle with the low signal-to-noise ratio of early-stage cancer VOCs, trained canines can produce a distinctive odor profile indicative of various cancer types.

SpotitEarly leverages this innate biological excellence through a meticulously designed system. Patients collect breath samples using an innovative and easy-to-use at-home breath collection kit that isolates VOCs and ensures their integrity during transit to the lab. These samples are then presented to qualified detection canines within a controlled laboratory environment. This “bio-sensing” step is critical, as it taps into a natural diagnostic ability that chemical and analytical sensors are far from being able to replicate.

Intelligence Unleashed: The LUCID AI Platform

While canine olfaction provides unmatched sensitivity, the challenge for large-scale clinical application lies in ensuring consistency, reproducibility, and the ability to process vast amounts of samples efficiently. This is where SpotitEarly’s LUCID platform comes into play, elevating the “bio” element with sophisticated AI technology.

The LUCID platform is SpotitEarly’s proprietary data and lab management system. It manages tests, stores data, and allows critical decisions throughout the diagnostic process, including determining the final result. The system integrates proprietary hardware devices, an array of sensors and software. It is designed to track samples from early registration to final diagnosis and automates the diagnostic process to minimize human intervention and reduce error rates.

Unmatched Accuracy through AI

The AI component within LUCID is crucial for achieving high diagnostic accuracy by amplifying the natural abilities of the detection canines. While dogs provide unparalleled sensitivity in detecting Volatile Organic Compounds (VOCs) associated with diseases, the AI adds precision and scalability.

Here’s how AI enhances accuracy:

• Data Integration and Interpretation: LUCID collects and analyzes thousands of data points every second, including physiological responses and behavioral cues from trained sniffer dogs as they screen samples. It also incorporates patient demographic and medical history information.
• Pattern Recognition and Refinement: Advanced deep learning algorithms analyze these diverse data streams to identify cancer-signal patterns. The AI learns continuously with each new dataset, improving its ability to distinguish true cancer signals from false positives.
• Confidence Scoring: The system provides a confidence score with each result output, quantifying the model’s confidence in its prediction. 
• High Performance Metrics: SpotitEarly’s clinical trials have demonstrated high overall accuracy (94.1%), sensitivity (93.9%), and specificity (94.2%) across four common cancer types (lung, breast, colorectal, and prostate). This performance includes impressive sensitivity of over 90% for early-stage detection (stages 1-2). 

Enabling Global Scale

AI is fundamental to the scalability of SpotitEarly’s solution, enabling high-throughput processing that would be impossible with canine-only methods.

• High Processing Capacity: A single lab facility equipped with the LUCID platform is designed to process over a million breath samples every year. This is achieved by optimizing dog performance and minimizing manual logistics through proprietary sniffing devices developed for this purpose.
• Global Replicability: The modular design of SpotitEarly’s labs and the LUCID platform ensures that facilities can be easily replicated worldwide. This focus on scalability also underpins the company’s commitment to equitable healthcare, ensuring life-saving screening can reach underserved populations regardless of geography or income.
• Cost-Effectiveness: By streamlining workflows through AI and canine olfaction, SpotitEarly reduces operational costs without compromising accuracy. The breath-based test is non-invasive and can be self-administered at home, further increasing convenience and accessibility.

A Paradigm Shift in Early Detection

SpotitEarly’s bio-AI hybrid system represents a disruptive approach to medical diagnostics. Current cancer screening methods can be expensive, invasive, and require point-of-care visits. Liquid biopsies, while less invasive, are currently less effective for detecting cancer at early stages. SpotitEarly’s technology addresses these limitations by developing a highly accurate, non-invasive, and cost-effective solution.

The collection process, using a breath collection mask, is designed for ease of use, allowing patients to perform the test at home or in clinical settings. This quick, non-invasive step minimizes friction with the healthcare system, benefiting underserved populations who might face barriers to traditional screening. The potential for saving lives and reducing healthcare costs is substantial; a 1% improvement in localized cancer diagnoses for lungs and bronchi alone could save 560 lives and $136 million annually. Applying this logic to other cancers could result in ten times the savings.

Competitive Differentiation

SpotitEarly’s bio-convergence model offers a pioneering and ambitious approach to healthcare diagnostics. By intelligently merging the biological strength of canines with advanced AI and data infrastructure, this model is designed to overcome many limitations of current screening methods. This fusion of natural and artificial intelligence promises a future where early disease detection is accurate, accessible, and transforms patient outcomes globally.

In the race to bring innovative medical products to market, startups in biotech and medtech face a paradox: they must move fast while navigating some of the most regulated, complex, and risk-sensitive environments in the business world. Traditional operational approaches are buckling under the weight of this dual pressure. Enter Artificial Intelligence, not as a buzzword, but as a pragmatic enabler of a new kind of operational intelligence.

Much of the hype around AI in healthcare revolves around diagnostics, imaging, and predictive modeling. Yet, behind every groundbreaking therapeutic lies a less glamorous but equally critical domain: operations. Supply chains, regulatory data management, and vendor qualification processes are riddled with bottlenecks that consume time, budgets, and human focus. Large Language Models (LLMs) and AI-native tools are now emerging as force multipliers, allowing lean teams to achieve operational resilience and compliance at scale.

Take supplier qualification as an example. A biotech startup aiming to produce a temperature-sensitive therapy may need to vet dozens of suppliers across packaging, logistics, and raw materials, each with its own regulatory documentation and audit requirements. Traditionally, this involves weeks of emails, PDF parsing, and Excel management. With AI-powered assistants trained on regulatory frameworks, startups can now automate the extraction, classification, and risk-scoring of supplier data, turning a multi-week process into a 48-hour sprint.

In one real-world scenario, a client in the pharma logistics space used a GPT-powered plugin built with OpenAI’s API and embedded in Google Workspace to analyze incoming supplier certifications directly from email attachments and generate a structured risk report that integrated with their QMS (Quality Management System). The tool saved over 60% of manual labor hours and eliminated redundant compliance checks.

Another area undergoing transformation is cold-chain logistics. Using predictive AI models, companies can anticipate where a shipment may be delayed or a temperature threshold exceeded, and trigger preemptive interventions. When LLMs are integrated into control tower interfaces, operations managers gain a real-time narrative summary of supply chain health, with plain-language suggestions and historical benchmarking, something previously reserved for Fortune 500 firms.

One biotech startup I supported built an internal GPT-based dashboard (not commercially available) using OpenAI’s API and integrated with Google Sheets and Slack that synthesized real-time shipment data from multiple carriers. The system provided alerts in natural language, such as “shipment 004 from Munich to TLV at risk of exceeding 8°C – ETA update suggests alternate routing via Prague,” empowering operations teams to respond without needing deep analytics skills.

Furthermore, the regulatory narrative, often seen as a burden,  is now becoming a strategic asset. AI can generate tailored audit trails, pre-populate technical files, and even draft QMS responses, ensuring that regulatory readiness is embedded into every phase of development. Of course, when implementing such solutions in such a sensitive field, data security and privacy must be paramount, with stringent security measures and adherence to strict privacy standards.

This is not about replacing humans. It’s about augmenting small, over-stretched teams with operational superpowers. In the coming years, startups that treat AI as part of their infrastructure, rather than an add-on, will be the ones that outmaneuver both incumbents and complexity. However, it’s crucial to acknowledge that despite the potential long-term savings, the initial investment costs in developing and implementing tailored AI systems can be significant, requiring careful budgetary planning. Furthermore, there are implementation challenges that need to be addressed, requiring technical expertise and the ability to integrate with existing systems.

But how can organizations trust AI tools in such sensitive domains?
Validation is a critical milestone.companies should evaluate these tools through sandbox testing (a way for trying out an AI tool in a safe, isolated environment), benchmark outputs against expert human review, and ensure they align with GxP or ISO-compliant documentation standards. Today’s leading tools increasingly provide audit trails, transparency dashboards, and explainability features that allow QA and compliance teams to trace each AI-driven decision. In regulated environments, it’s not just about speed it’s about reproducibility, traceability, and risk mitigation.

Example of tools supporting validation and auditability:

• TruLens: Tracks LLM behavior, provides explainability and evaluation of AI responses. trulens.org
• Weights & Biases: Monitors AI model runs, parameters, and outcomes for full traceability. wandb.ai
• ClearML: Offers reproducibility, experiment versioning, and audit logging across the ML lifecycle. clear.ml

The operational backbone of innovation is being rebuilt, not with more people, but with smarter systems. And for medtech and biotech startups, that shift is not a luxury, it’s the edge, a winning position !

Artificial Intelligence (AI) has made remarkable progress in recent years, offering promising tools to streamline documentation, accelerate data analysis, and support planning as well as strategic and regulatory decision-making across the product development lifecycle. However, when applied to regulatory writing and scientific interpretation, especially in the preparation of regulatory development plans and formal submissions such as Pre-RFDs, Pre-INDs, or Scientific Advice packages, current AI tools reveal significant limitations. These shortcomings pose meaningful challenges for developers of drugs, medical devices, and combination products, potentially resulting in regulatory communication gaps, misclassification, or flawed strategic decisions that can result in substantial delays, increased resource expenditure, and an extended time-to-market.

Misalignment with Regulatory Language and Strategic Intent

One of the most significant challenges of AI-generated content is its frequent misalignment with the precise and context-sensitive language required in regulatory communication. While AI tools can produce fluent, grammatically correct English, they often distort the intended regulatory message in subtle but meaningful ways.

For instance, when drafting a Pre-RFD to support the classification of a product as a medical device, AI may introduce terminology commonly associated with pharmaceutical products. What may appear as minor linguistic choices, such as referring to “active ingredients,” “systemic effects,” or “pharmacological action”, can conflict with the regulatory requirements for devices. This is particularly critical when describing the product’s mechanism of action, which must not only align with regulatory definitions of medical devices but also consider the diverse classification frameworks and terminological nuances applied by health authorities across different jurisdictions.

Inaccurate language may suggest pharmacologic activity where none exists, potentially triggering misclassification, increased regulatory hurdle, or delays in review. Moreover, given the variability in terminology and classification criteria across jurisdictions, regulatory messaging must be carefully tailored to each specific context, something current AI systems are not reliably equipped to do.

In pursuit of sounding more polished or “native,” AI tools also tend to replace specific regulatory terminology with broader or stylistically refined alternatives. This can compromise the scientific clarity and regulatory intent of a submission, which may significantly impact regulatory interpretation and decision-making.

AI tools are not yet capable of reliably interpreting nuanced regulatory distinctions or adjusting language to support the strategic regulatory positioning of a product effectively.

Challenges in Clinical Data Retrieval and Interpretation

AI tools are increasingly used to assist in identifying and analyzing large databases such as clinical trials from public registries and other platforms. However, their ability to retrieve specific studies or datasets, mainly when based on unique identifiers like NCT numbers, is still limited. In many instances, AI-generated outputs return incomplete results, overlook key endpoints, or misrepresent clinical aspects of study design and findings. These inaccuracies may stem from limitations in recognizing trial identifiers, differentiating between product classifications, and other formal definitions.

Beyond these technical limitations, a more fundamental challenge lies in AI’s inability to contextualize clinical data within the specific development stage of a product. For example, AI-generated analysis may fail to recognize whether the product has already undergone safety evaluations in previous studies, whether it is approved and now being studied for a new indication, or whether it is a novel investigational product. These distinctions are critical for assessing the relevance, novelty, and regulatory interpretation of the data.

In addition, AI tools generally do not account for broader clinical and methodological context—such as how the selection of primary and secondary endpoints aligns with the study’s inclusion and exclusion criteria, how these endpoints relate to the overall study duration and follow-up period, or whether the analysis focuses on a single timepoint versus longitudinal data.

As a result, the evidence summaries produced by AI may misrepresent the maturity or adequacy of the clinical dataset. When such outputs are used to inform development strategies or formal regulatory submissions, they can lead to misguided clinical assumptions, suboptimal protocol designs, inefficient prioritization of studies and milestones, and ultimately fail to align with regulatory expectations.

Inaccurate or Incomplete Referencing of Scientific Literature

Sourcing and citing peer-reviewed literature is another common area where AI tools fall short. When prompted to retrieve articles using DOI numbers or extract references from a predefined literature list, AI tools often fail to align citations with appropriate content, returning entirely incorrect sources, or, in some cases, fail to retrieve any results at all.

Even more concerning is the use of AI to generate scientific content intended to support regulatory submissions, where tools have been known to fabricate citations entirely. This not only undermines the scientific integrity of the document but also poses a significant risk to the credibility of the submission if unverifiable or non-existent references are included.

The Future of AI in Regulatory Planning: Progress with Caution

AI holds considerable promise as a supportive tool in the regulatory processes surrounding drug and medical device development. It can be a powerful assistant for early-stage drafting, language refinement, and high-level summarization. Additionally, AI has the potential to save time and resources when analyzing large datasets, helping to inform more robust regulatory assessments and support the strategic design of development plans.

As AI tools continue to advance, several current limitations in regulatory writing and data assessment may become less prominent. Structured and harmonized data environments, combined with enhanced natural language understanding models, may allow AI systems to more consistently extract relevant information, and tag key endpoints. This will reduce the manual effort involved in basic data mining and speed up early-stage analysis.

However, despite these gains, one of the most persistent and problematic gaps will remain: the inability to independently verify the accuracy or validity of such AI-driven analyses. Even if AI systems can surface studies based on seemingly correct filters or terminology, there is currently no mechanism to audit or validate how these tools weigh relevance, detect bias, or infer conclusions from aggregated data. AI lacks epistemic awareness: it does not “know” when it’s wrong, nor can it justify its outputs with the same methodological transparency required in regulatory contexts. As a result, developers may still face a critical verification burden when using AI-derived evidence to support clinical assumptions or regulatory arguments.

At its current level of maturity, AI cannot replace the expertise of regulatory professionals, especially when precision, context sensitivity, and the articulation of a clear clinical and regulatory strategy are critical to the product development plan and overall regulatory success. Organizations developing drugs, devices, or combination products should remain cautious when leveraging AI for regulatory purposes. Developers relying on AI-generated text, regulatory assessment, or clinical designs without expert oversight and integration of product-specific knowledge risk undermining their own classification strategy and introducing avoidable regulatory hurdles. Until these technologies evolve to fully comprehend regulatory frameworks, classification pathways, and the complexity and regulatory significance of formal submissions, their role should remain advisory and supplementary, rather than serving as a primary decision-making tool.

The advent of new technology always ushers increasingly complex developments in the ever-evolving landscape of drug development. The uptake of Artificial Intelligence (AI) technologies has been ubiquitous in all areas of drug development, including clinical research where digital health solutions are being employed to increase clinical trial efficiency and decrease the associated time and costs.

Clinical trials are fraught with the resource-intensive hurdles of cost, time, and complexity. A promising application of AI being used to address these issues is digital twins. Digital twins are digital replicas of physical objects or systems connected by bidirectional data and information flow. Popular in the aerospace and manufacturing industries, digital twins are also being used in clinical trials to replicate biological systems or processes to simulate real time biological processes and to model outcomes.

Digital twins can model biological components ranging from cells and tissues to organs and environments in a patient’s body. A digital twin is generated from preexisting data, AI modeling and incorporates real time data to predict outcomes to optimize decision making. These twins are versatile and have several applications, some of which include drug discovery, drug repositioning, personalized treatments based on digital patient profiles, recruitment into trials as virtual patients, in-silico clinical trial design and safety monitoring.

Featured are a small variety of companies demonstrating the creative applications of digital twin technology in clinical trials:

• Unlearn –  Unlearn has a platform to generate digital twins aiming to aid in designing more efficient trials, reducing sample sizes, boosting power, and making faster, more confident development decisions. PROCOVA™ is a statistical methodology developed by Unlearn.AI for incorporating prognostic scores derived from trial participants’ digital twins into the design and analysis of phase 2 and 3 clinical trials.

This methodology has been qualified by the EMA and is covered under the FDA’s guidance on Adjusting for Covariates in Randomized Clinical Trials for Drugs and Biological Products as a special case of ANCOVA statistical method.

• BOTdesign –  Botdesign has ORIGA, Europe’s first web-based platform for augmenting clinical data with deep learning. It enables healthcare manufacturers and researchers to generate realistic artificial patients, while guaranteeing data confidentiality and regulatory compliance. ORIGA is based on advanced generative AI models called Variational Autoencoders (VAEs) used to create synthetic patients. This can be particularly useful in increasing size and diversity in research especially for rare and underrepresented cohorts.
• Aitia –  Aitia has built a causal AI engine (REFS®) that uses high-performance computational power to turn massive amounts of multiomic and patient outcome data into fully realized, unbiased and causal in silico models of human disease called “Gemini Digital Twins” that can be used to discover new causal human drug targets and biomarkers, candidate patient subpopulations for clinical trials, and optimal drug combinations.
• Bayer – Bayer has used digital twins to create virtual trial arms or “external control arms”, which can replace control/placebo arms in some clinical trials. This can help fill evidence gaps e.g., where an RCT (randomized control trial) is not feasible or ethically sound, in addition to reducing costs, overall development time and/or trial recruitment time.
• Sanofi – Sanofi uses quantitative systems pharmacology (QSP) modeling of a disease and available clinical trial data from live patients to create digital twins of the human patients seen in the clinic. All of the available data on disease biology, pathophysiology, and known pharmacology, is taken and integrated into a single computational framework.

Although digital twins can’t fully substitute real humans, they can help streamline clinical trials by reducing costs and timelines. As is the case with any technology there are associated ethical, technological and regulatory risks and challenges. The accuracy and predictive power of a digital twin heavily depend on the quality of input data, and issues with generalizability currently limit scalability. Given their extensive reliance on patient data, digital twins must comply with the varied privacy and security laws globally. Nevertheless, the advancement of AI technologies lends potential for digital twins to revolutionize drug discovery and development even further.

Placing an AI-powered medical device on the EU market requires complex strategies and a high level of both technical and regulatory expertise. This even gets trickier for Software as a Medical Device (SaMD) powered by Artificial Intelligence (AI).

Choosing the Right Notified Body

Notified Bodies are responsible for evaluating medical devices in accordance with EU MDR requirements and issuing the CE Certificate. One of the first and most critical steps is selecting a Notified Body with reviewers who have sufficient expertise in AI technologies. Without this, the conformity assessment process can become inefficient or misguided.

Classification Challenges

One of the most common issues with AI-powered SAMDs lies in their classification. At this point, a clear idea about the real clinical benefit of the device is the most important issue. In most cases, the pathway to getting clearance goes through this clarification. Manufacturers frequently either overstate or understate the clinical benefit. This benefit, for example, can be decreasing the time spent by a clinician during a routine procedure or directly marking the tumor. Deciding on which would change its classification.

Demonstrating Clinical Benefit

Once clearly defined, it must be supported by robust evidence. Typically, this is achieved through collecting and analyzing clinical data. However, it shall be noted that traditional methodologies used for a physical medical device often do not fit the software domain well. The intended clinical benefit will also have major effect on the necessity of prospective studies.

Statistical Validation

Statistical success is the key aspect of validating the AI tool. Therefore, manufacturers should be very careful about selecting the most suitable statistical methods and study design. It is essential to consider multiple dimensions of performance, such as sensitivity, specificity, accuracy, etc.

Software and AI Development Lifecycle

Notified Bodies depend on a software development lifecycle as much as their validation results when approving software as a medical device. Therefore, one key aspect of success is integrating the AI Development Lifecycle into the existing software development lifecycle. This includes security management, AI Data Management, Data Labeling Practices, the AI Model Development Phase, Procedures for training, evaluating, and documenting AI models, as well as their release and maintenance.

While harmonized standards offer general guidance, they may not fully reflect the state of the art for AI systems. Manufacturers are encouraged to review frameworks such as CMMI for more mature development process insight.

Stay Up to Date

Manufacturers should maintain constant awareness of:

• EU Regulations
• Relevant MDCG guidance documents
• Applicable EN/ISO/IEC standards
• Team-NB Documents
• IG-NB Documents

These documents evolve and often contain critical clarifications that directly impact AI-based SaMD development and approval.

Artificial intelligence is becoming an essential component of modern clinical trials. It supports patient recruitment, accelerates data analysis, enables adaptive trial designs, and contributes to regulatory decision-making. As sponsors adopt AI systems across various stages of the research lifecycle, they must address the legal and ethical frameworks that govern the use of personal data and algorithmic technologies in healthcare.
This article outlines the core responsibilities of clinical trial sponsors when using AI tools, with a primary focus on European data protection and AI regulations, while also referencing global guidance and emerging standards that shape the broader landscape.

Applying the General Data Protection Regulation

The General Data Protection Regulation applies whenever personal data is processed in the European Union (“EU”), including the European Economic Area (“EEA”) and the United Kingdom (“UK”), or by entities targeting individuals in the EU. When an AI system is used in a clinical trial to process personal data, the sponsor qualifies as a data controller and remains responsible for GDPR compliance of all data processing activities. This includes verifying the legal basis for processing; typically consent or legitimate interest of the sponsor in the area of scientific research; ensuring that the AI system operates in line with the purpose limitation principle and drafting the necessary records and assessments to demonstrate accountability. If an AI tool is introduced after initial data collection, or if its function differs from that initially communicated, the sponsor may need to assess whether the original legal basis still applies or comply with the obligations required for establishing a new legal basis. The sponsor must also ensure that any third-party providing AI services operates as a processor under a compliant data processing agreement and implements adequate technical and organisational measures to protect data confidentiality and integrity.

Understanding the EU AI Act and Its Interception with GDPR

In 2024, the European Union adopted the AI Act, which establishes a legal framework for the development and use of artificial intelligence systems. The regulation applies to all AI systems that are placed on the EU market or used in the EU, regardless of where the provider is based.
The EU AI Act establishes a risk-based regulatory framework that classifies AI systems into four categories: unacceptable, high, limited, and minimal risk. Unacceptable-risk systems are banned outright within the EU because they pose a serious threat to fundamental rights, safety, or democratic values. High-risk AI systems are subject to the most stringent obligations under the Act and can only be deployed if in compliance with such obligations. Limited-risk AI systems are permitted but must meet basic transparency requirements, such as informing users they are interacting with an AI system and ensuring the design does not mislead or deceive. Minimal-risk systems are not subject to specific requirements under the AI Act, but must still comply with other applicable laws, including data protection frameworks.

It is important to clarify the difference between the AI and Data Protection legal frameworks and why they may work simultaneously. While the GDPR applies to the processing of personal data (and if an AI system processes personal data, the person responsible for deploying such system must comply with the GDPR) the AI Act serves to ensure the ethical use of AI in real-world situations. Does the use of AI always require the processing of personal data? Not necessarily. However, in clinical trials, most AI applications, such as patient selection, imaging analysis, and safety monitoring, typically involve personal data.

Does the EU AI Act Apply to Scientific Research?

Despite the broad scope of the AI Act, Article 2(6) and Recital 25 establish a narrow exclusion for AI systems developed and used exclusively for scientific research and development. According to the Regulation, such systems fall outside the AI Act only if they are created solely for the purpose of conducting scientific research, and only if they are not placed on the market or used to produce legal or significant effects on individuals.

This exclusion was introduced to protect academic and experimental research and is designed to avoid imposing the full regulatory burden on AI models used in non-commercial, closed research environments. However, the exemption does not apply in a number of common clinical research scenarios. First, if the AI system is procured or licensed from a commercial provider, rather than developed specifically for the research project, the exclusion cannot be claimed. Second, if the system is used in a clinical trial where it influences patient eligibility, dosing, safety monitoring, or any aspect of the investigational product’s development pathway, the system is no longer considered confined to a purely scientific function. It is then considered to be “put into service,” as defined in Article 3(13) of the AI Act.

In practice, this means that most AI tools used operationally in clinical trials, particularly in interventional or regulatory-driven settings, will not qualify for the scientific research exclusion. The same applies to systems developed in a research environment but intended for future market use, including tools supporting software as a medical device or algorithms subject to future certification.

EU AI Act and Clinical Trials

AI systems used in clinical trials may fall within the high-risk category under the EU AI Act through two regulatory pathways outlined in Article 6. First, under Article 6(1), an AI system is considered high risk if it is a product or a safety component of a product governed by EU harmonization legislation listed in Annex I, such as medical devices under Regulation (EU) 2017/745 or in vitro diagnostic devices under Regulation (EU) 2017/746, and if that product requires third-party conformity assessment. This means that investigational AI tools used for diagnostic decision support, patient stratification based on biomarkers, or real-time safety monitoring may be classified as high risk if they fall within the scope of these device regulations and are subject to notified body review.

Second, Article 6(2) states that AI systems listed in Annex III are also deemed high risk. While clinical research is not explicitly mentioned in Annex III, an AI system used in a trial may fall under this category if it materially influences decisions that affect participants’ health or fundamental rights, particularly where profiling is involved or medical decision-making is impacted. Sponsors must assess whether the AI system qualifies under either of these routes, as both may lead to a high-risk designation with corresponding regulatory obligations.

If a clinical trial sponsor deploys a high-risk AI system (e.g. for patient selection, safety signal detection, or diagnostic support), it must comply with the EU AI Act by ensuring the system is used according to the provider’s instructions, assigning trained human oversight, retaining system logs for at least six months, and monitoring the system’s performance. The sponsor must report any serious incidents or risks to the provider and relevant authorities without delay, ensure input data is relevant and representative, inform trial participants of the AI system’s use, and where applicable, perform a fundamental rights impact assessment and complement the existing GDPR Data Protection Impact Assessment (DPIA) with AI-specific risks.

The Role of Data Protection Impact Assessments

When AI systems are used in clinical trials and involve the processing of sensitive health data or automated decision-making, a Data Protection Impact Assessment may be required under the GDPR. This assessment should include a description of the processing, the purpose of the AI system, the legal basis for data use, and an evaluation of the risks to data subjects. Where the AI system falls under the AI Act’s high-risk category, the sponsor must also maintain a risk management framework aligned with the requirements of the Regulation, including appropriate levels of human involvement, accuracy monitoring, and transparency in system design.

Global Context: Ethics and Emerging Regulatory Approaches

While the European Union provides one of the most comprehensive legal frameworks for AI in healthcare, other jurisdictions are developing their own regulatory and ethical approaches. The United States Food and Drug Administration (FDA) has issued an action plan for AI in medical devices and emphasizes good machine learning practices, particularly in software that evolves over time. Health Canada has issued draft guidance for AI-enabled medical devices, and Australia has adopted a regulatory sandbox model for early-stage AI testing.

The World Health Organization has published the Ethics and Governance of Artificial Intelligence for Health report, which sets out guiding principles such as transparency, accountability, inclusiveness, and respect for human autonomy. These principles are intended to guide all stakeholders involved in health-related AI, including researchers and sponsors. Even where specific legal obligations may not yet exist, adherence to ethical standards is increasingly expected by ethics committees, funders, and regulatory agencies. Sponsors are encouraged to align with these international standards and document their governance processes accordingly.

Conclusion

The application of the EU AI Act follows a phased approach. The Regulation entered into force in August 2024, with key provisions becoming applicable in stages. Rules concerning prohibited AI practices and AI literacy take effect from February 2025. Obligations for general-purpose AI systems, including transparency, documentation, and risk mitigation, will apply from August 2025. Requirements for high-risk AI systems, such as conformity assessments, risk management, and human oversight, come into force from August 2026. For AI systems embedded in medical devices that require notified body involvement, the relevant obligations apply from August 2027.

At the same time, jurisdictions such as the United States, Canada, the United Kingdom, and Australia are developing or implementing new legal frameworks to govern the use of AI in healthcare and clinical research. As global standards continue to emerge, clinical trial sponsors should design compliance programs that align with both European regulations and international expectations. A harmonized approach will help ensure ethical, legal, and operational consistency when deploying AI tools in trials across multiple regions.

The convergence of artificial intelligence (AI) and organoid technologies is beginning to reconfigure the early stages of drug development. These two innovation domains, each advancing rapidly on their own, are now intersecting in ways that promise to improve the predictive value of preclinical testing, reduce the cost and duration of development pipelines, and ultimately produce safer, more effective therapies. Yet alongside this opportunity lies a complex set of technical, ethical, and regulatory challenges. For the scientific and biotech community, navigating this evolving landscape will require not only technological adaptation but also institutional coordination and policy foresight.

A New Convergence in Preclinical Modeling

Organoids – three-dimensional, multicellular constructs derived from stem cells – have emerged as biologically relevant in vitro systems that recapitulate aspects of human tissue architecture and function. Their ability to model complex human phenotypes has led to growing use in oncology, infectious disease, toxicology, and regenerative medicine. Compared to animal models or two-dimensional cultures, organoids offer advantages in terms of genetic fidelity, species relevance, and personalization. However, their adoption in industrial drug pipelines remains limited by variability in culture protocols, inconsistencies in functional readouts, and a lack of data harmonization across producers and laboratories. These limitations have motivated increasing interest in computational approaches to standardize interpretation and enhance comparability – enter AI.

Machine learning and deep learning approaches, when applied to the outputs of organoid systems, can extract latent patterns in high-dimensional data, such as transcriptomics, high-content imaging, and pharmacological response profiles. AI has shown promise in identifying phenotypic signatures, classifying tissue states, and predicting drug responses. In theory, these tools could accelerate compound screening and guide mechanism-informed lead selection. Yet AI systems trained on organoid data inherit the uncertainties and inconsistencies of their biological source material. As a result, successful integration depends on improving both experimental standardization and data quality—two prerequisites for effective model training, validation, and interpretation.

Regulatory Realignments and the Burden of Proof

The regulatory environment is evolving in parallel. In the United States, the passage of the FDA Modernization Act 2.0 in 2022 formally removed the requirement for animal testing prior to human trials. This shift has created space for new approach methodologies (NAMs), including organoids, computational simulations, and other alternatives, to support investigational new drug (IND) applications. The FDA’s Model-Informed Drug Development (MIDD) initiative encourages the use of simulation and predictive modeling throughout the development process. Simultaneously, the agency has begun developing frameworks for AI/ML-based software, focusing on algorithmic transparency, real-world validation, and risk mitigation. While regulatory acceptance of AI-derived predictions remains cautious, the direction is clear: tools that are well-characterized, traceable, and biologically grounded are increasingly welcome in preclinical and regulatory workflows.

In the European Union, a more prescriptive and comprehensive regulatory framework is emerging. The Artificial Intelligence Act, adopted in 2024 and set to be enforced in phases from 2025, represents the first region-wide legislation governing AI. Biomedical applications—particularly those with potential implications for health outcomes—are designated as “high-risk” under the Act. Developers must meet requirements related to data governance, explainability, human oversight, and post-market monitoring. Although the Act is technology-neutral, its implications for AI-driven drug development are significant, especially when organoid-derived or patient-specific data are involved. Unlike sector-specific guidance, the AI Act applies horizontally across domains, which presents both a compliance burden and an opportunity to build AI tools that are safe, auditable, and trustworthy by design.

Toward a Predictive and Accountable Innovation Ecosystem

Despite their promise, the integration of organoids and AI into drug development raises systemic challenges. A persistent lack of protocol standardization continues to limit reproducibility across labs and platforms. Biological heterogeneity, while valuable for capturing patient diversity, also complicates benchmarking and model generalization. The ethical use of patient-derived tissues and associated data requires robust consent procedures and governance structures that can support both research and commercial applications. On the computational side, many AI models function as black boxes, limiting interpretability and regulatory acceptability. Moreover, the successful deployment of these technologies depends on interdisciplinary teams—yet the integration of wet-lab biology, computational modeling, and regulatory expertise remains rare in most research environments.

Nevertheless, a growing body of academic, industry, and regulatory stakeholders is working to address these gaps. Efforts to create interoperable organoid databases, define reference standards, and foster precompetitive data-sharing frameworks are underway. Some regulatory agencies are exploring sandbox initiatives that allow developers to test AI models in controlled settings with early feedback. Ethical frameworks for the secondary use of patient-derived data in AI training are also gaining attention, although global harmonization remains limited.

In the years ahead, the integration of AI and organoid platforms could enable a more human-relevant and predictive approach to drug development—one in which computational models are trained on real biological complexity, and preclinical decisions are informed by tissue-specific responses. But realizing this potential will require more than innovation. It will demand transparency, shared standards, and a long-term commitment to collaborative infrastructure. The scientific community must work not only on the frontiers of technology, but also at the interface of governance, ethics, and reproducibility.

In this context, the convergence of AI and organoid science is not simply a technical advance. It is a shift in how we conceptualize preclinical research—away from generalized proxies and toward systems that integrate human biology, computation, and regulatory science in a coherent, scalable, and accountable way.