Review of the FDA Draft Guidance on AI in Drug Development

1. For those less familiar with the technology, how would you describe the difference between AI and ML in simple terms, and why does that distinction matter in pharma?

Artificial intelligence (AI) is the general idea of computers doing things that normally require human intelligence, like identifying patterns, making predictions, or aiding decisions. Machine learning (ML) is a subset of AI that learns from data instead of just relying on predefined rules.

From a pharma perspective, the distinction is less important than how the technology is deployed. One thing I like about the draft guidance is that it shifts the conversation away from whether something is called AI or ML and focuses instead on whether the model is fit for purpose and supported by appropriate evidence.

What really counts is the model’s context of use, its impact on decisions, the possible consequences of mistakes, and whether the proper controls are in place to safeguard patient safety and product quality.

 

2. What are some real-world examples of how AI is already being used successfully in drug development today?

We are already seeing AI create value across the pharmaceutical lifecycle, particularly in areas where there’s a large volume of complex data. From drug discovery and clinical development to manufacturing, quality and pharmacovigilance, AI helps scientists identify patterns, make predictions and support decision-making.

For example, in non-clinical development, AI is being used to predict toxicity, identify off-target effects, and prioritize promising compounds. In clinical development, it’s helping improve patient stratification, optimize trial design, support recruitment strategies and analyze complex clinical datasets.

On the manufacturing and quality side, AI is being applied to process monitoring, predictive maintenance, yield optimization, deviation classification, and document review. It’s also increasingly being used in pharmacovigilance to support adverse event case triage and safety signal detection.

What they all have in common is that AI primarily serves as a decision-support tool. It complements scientific expertise and helps people make better, faster decisions rather than replacing human judgment.

 

3. Your talk at the Summer Scientific Forum discusses the new FDA draft guidance on AI in drug and biological product development. What are the most important takeaways for pharma companies?

The most important takeaway is that the FDA is providing a practical, risk-based framework for evaluating AI credibility.

The guidance encourages organizations to start with a clear definition of the question of interest and the context of use. From there, they should assess model risk based on model influence and decision consequence.

Companies should then establish credibility goals, generate relevant evidence and ensure oversight throughout the model lifecycle.

Another important message is that there is no one-size-fits-all validation strategy. The level of evidence required should be proportional to the model risk and intended use.

Finally, the guidance highlights the sponsor’s responsibility. No matter how sophisticated an AI model may become, the ultimate responsibility for patient safety, product quality and regulatory compliance rests with people. I actually see the guidance as an enabler. It gives companies a practical framework so they can adopt AI confidently while still meeting expectations for quality and patient safety.

 

4. The FDA guidance is still in draft form, what aspects do you think are most likely to change, and what are you hoping they’ll clarify or modify based on industry feedback?

Overall, I think the direction of the guidance is solid, and I don’t expect the core principles of risk-based oversight, transparency, validation, lifecycle management and human accountability to change significantly. These concepts are already well aligned with existing regulatory expectations for quality systems, analytical methods and computerized systems.

What I’m looking for is more clarity on how the guidance will be implemented in practice. One area is the level of documentation and evidence required for different categories of AI applications.

Another area is lifecycle management for adaptive or continuously learning models. While the guidance appropriately emphasizes ongoing monitoring and revalidation, sponsors would welcome greater clarity on when model updates would trigger regulatory interactions and on the level of revalidation expected following model changes.

I also believe the industry is looking for additional guidance around data quality, bias assessment, and the use of third-party or commercially developed AI models. As AI ecosystems become more complex, companies require clarity on accountability when models, data sources or underlying technologies are provided by external providers.

Overall, I hope the final guidance preserves its flexible, risk-based framework while providing greater clarity on implementation.

 

5. You mentioned AI applications across discovery, clinical development, manufacturing, quality and pharmacovigilance—which area do you think offers the most immediate value?

AI is creating value across the pharmaceutical lifecycle, but I believe the most immediate value today lies in clinical development, simply because this is where the industry faces some of its greatest challenges in cost, timelines and complexity.

Clinical trials generate enormous amounts of data and often account for the largest portion of development costs. AI can also help improve patient recruitment, improve protocol design, find suitable trial sites, and analyze more complex clinical data.

That said, from an operational perspective, manufacturing and quality may offer the fastest and most measurable return on investment. AI can improve process monitoring, predict equipment failures, detect process deviations sooner, reduce batch failures and allow more reliable product quality. These applications often use existing manufacturing data and can deliver tangible business benefits, while operating within well-established quality frameworks.

So, if I had to choose one area for immediate impact, I would say clinical development because of its direct impact on development speed, cost and probability of success.

 

6. Could you share some strategies to mitigate bias and model drift?

Bias mitigation and model drift are among the most important considerations when deploying AI in regulated pharmaceutical environments. Even a highly accurate model can become unreliable if underlying data or operating conditions change over time. It really starts with good data. If your training data is not representative or well governed, you’ll introduce bias before the model is even deployed. It’s crucial to know the strengths and limitations of training data. To fight drift, organizations need to develop continuous performance monitoring programs, establish meaningful performance metrics, and investigate unexpected shifts in model behavior. If performance degradation is observed, there should be predefined processes for reassessment, retraining, revalidation and change management.

The most important factor is that mitigation activities should be commensurate with model risk, with higher risk applications requiring more robust monitoring, documentation, oversight and lifecycle controls. Ultimately, the most efficient plan is to treat AI models like any other GxP-relevant system: implement robust data governance, validate the model for its intended use, monitor its performance continuously, and have appropriate human monitoring throughout the model lifecycle.

 

7. What resources would you recommend for people who want to learn more about AI in pharmaceutical development?

For professionals that are interested in AI in pharmaceutical development, I would suggest starting with regulatory and industry resources, rather than focusing only on the technology itself. Knowing the regulatory context is essential because successful AI adoption in pharma is as much about governance, validation and quality systems as it is about algorithms.

I would start with the FDA’s draft guidance because it explains the agency’s current thinking. After that, the FDA discussion paper on AI in drug development provides useful background. I also recommend looking at Good Machine Learning Practice principles and the relevant ICH quality and risk management guidelines. Beyond regulatory guidance, organizations like the International Society for Pharmaceutical Engineering (ISPE), the Drug Information Association (DIA), and other professional associations regularly publish practical case studies and implementation experiences that show how companies are successfully applying AI across the pharmaceutical industry.