A conversation with Paul Hanson of Takeda Pharmaceutical
The real potential of generative artificial intelligence in pharmaceuticals may be similar to its contributions to other fields: not so much amazing breakthroughs, but reducing the amount of work humans do by offloading some of the mundane tasks to computers.
Paul Hanson, head of lifecycle management, innovation and strategy at Takeda and an active member of the International Society for Pharmaceutical Engineering (ISPE), made the case for GenAI's usefulness as a helper at the 2024 ISPE Europe Annual Conference earlier this year.
In his talk, titled “How Generative AI Will Impact Pharmaceutical Manufacturing,” he used the phrase “frictionless access,” a phrase that will resonate with anyone who has ever wandered through a file directory looking for a document.
We caught up with Hanson after the conference to get the rundown, and here's what he told us:
You gave the example of the steel industry using generative AI to help technicians complete work orders faster. How can this be applied to the manufacturing of complex molecules?
Hanson: US Steel provides its maintenance teams with transparency and seamless access to documentation related to their daily work, and we see a similar opportunity in complex molecular manufacturing.
One example is providing frontline workers with seamless access to process knowledge, enabling them to quickly resolve process-related deviations.
You covered some of the big risks in your talk: data leaks, flawed outputs, and the resulting loss of public trust if AI systems fail badly. What are some mitigation strategies to avoid these? Will we still need human oversight to double-check everything?
Hanson: There are several strategies to mitigate these risks. The first is quality: it is important to ensure that the data used by the model and presented to the end user meets data integrity standards.
Then, apply appropriate levels of control throughout development, certification, and post-launch lifecycle management to ensure the models are suitable for their intended use. Finally, methods such as search expansion generation allow companies to apply GenAI to structured and unstructured data separately, minimizing the leakage of proprietary information.
As for the need for human oversight, currently the relationship between humans and machine learning must be governed by the risk of how the model's recommendations might affect patients.
As our quality systems co-evolve with these models (e.g., continuous process validation), the number of human touch points will decrease, but never go to zero. This is a good outcome because, at the end of the day, we, as manufacturers, are connected directly to patients through the decisions that ensure the safety, purity, and efficacy of our products. This connection with patients reinforces our commitment to delivering high-quality products and the processes that support them.
Do you believe the potential benefits outweigh the risks? Please explain.
Hanson: The answer to this question is based on the last word: risk. In this context, quality maturity models become dramatically more relevant for enterprises. Specifically, maturity becomes the differentiator that ensures the success of these machine learning models (e.g., knowledge management).
As an organization progresses along the path of the Quality Maturity Model, the friction in achieving the value proposition of these GenAI models will naturally decrease.
One of the reasons for the growing need for these systems in the manufacturing of complex molecules is the fact that roughly 30 million workers will retire in the next six or so years. The number of workers entering the system will not be sufficient to replace them, so the industry needs tools to fill the knowledge gap. GenAI is one tool that can help structure knowledge for the next generation of workers, who, incidentally, are already using these knowledge-based models frequently.
How does the Quality by Design methodology fit into the AI discussion?
Hanson: We think translating quality by design into GenAI is very exciting, because QbD is a control framework fundamentally understood by health authorities. Translating QbD into GenAI means starting from the original purpose of the machine learning model and characterizing how different inputs (structured and unstructured data) affect its performance.
Once in production, the model's performance is further characterized to ensure it remains as expected as the associated manufacturing knowledge base grows. If performance deviates from expectations, the underlying model needs to be characterized more deeply to understand how the increasing scale of the inputs is affecting performance. Once performance returns to expectations, the updated model can be released for broader adoption within the organization. This virtuous cycle leads to a better understanding and explainability of the model's performance, defining the “operating space” of knowledge for users.
Can you outline some of the first small steps that biotech companies can take to introduce AI into their processes?
Hanson: Start by defining a small, low-risk, and specific use case that should produce tangible results. Then characterize GenAI's performance within the workflow of your use case. For example, if you're using GenAI to summarize technical reports, assign different personas to GenAI and request different levels of detail to understand results across different domains (accuracy, readability, etc.).
We then use the results to establish what we call the jagged boundary of knowledge. The shape of this boundary ultimately determines the boundary of trust we give to the model's features. We then periodically publish the results to our user base so that they can stay updated on the limits of their confidence in the model. Finally, we allow users to provide feedback about their experience. User voice is critical to understanding experiences with our models at scale and minimizing the risk of blind spots when it comes to system performance.
About the experts:
Paul Hanson has been Head of Lifecycle Management, Innovation and Strategy at Takeda Pharmaceutical since 2007 and is an active member of ISPE. He began his career in the R&D Bioprocess Development group, moving from late stage assets to marketed products. Paul moved with those products to lead the newly formed Biologics group in commercial technical operations. With Takeda's acquisition of Shire in 2019, Paul took on a new role as Head of Lifecycle Management, Innovation and Strategy within the recently formed Global Manufacturing Science organization. In this role, he leads a diverse group working in the areas of global knowledge management including materials qualification, dedicated manufacturing investigators and more recently the qualification and application of artificial intelligence in manufacturing business processes.
