Artificial intelligence and machine learning (AI/ML) are breakthrough tools in biopharmaceutical processing, but they are only as good as the dynamic processes they interact with. You know the “garbage-in, garbage out” adage, but manufacturers need to look beyond those obvious inputs and ensure the entire bioprocessing operation is designed for sustainable AI/ML applications within the regulatory framework.
“AI/ML models can offer great benefits to the industry when implemented using a risk-based approach,” researcher Gowtham Nakka, University of Louisiana at Lafayette, tells GEN. With that in mind, “Regulatory agencies have to do more to provide practical guidance to the industry.”
There are three technical enablers “that repeatedly separate successful [AI] deployments from stalled pilots: trustworthy data pipelines, fit-for-purpose model selection, and lifestyle governance,” according to a paper by Nakka, university colleague Katti Kartik Reddy, and independent researcher Sakshi Gupta.
Trustworthy data, the first of those enablers, is ensured by “data monitoring systems that provide thorough tracking with complete traceability,” they point out. “Data pipeline maturity includes time synchronization, consistent tag naming, audit trails, and integration of material genealogy (lots, suppliers) with process and/or lab data. Without these, models may appear to work in retrospective analyses but fail when deployed.”
As the second enabler, fit-for-purpose models, the authors highlight the Dynamic Mode Decomposition with Control (DMDc) model introduced in 2025 by Consuelo Vega-Zambrano, PhD student, and Vassilis M. Charitopoulos, PhD, associate professor, department of chemical engineering, both of University College London, and Nikolaos A. Diangelakis, PhD, assistant professor, school of chemical and environmental engineering, Technical University of Crete.
DMDc is “an explainable machine learning method to model continuous manufacturing in the pharmaceutical industry,” Charitopoulos’ team explains. It captures nonlinear dynamics of multiple input and output systems, thereby improving pharmaceutical continuous manufacturing modeling and control “without hindering trust.”
AI lifecycle governance
The third enabler, AI lifecycle governance, is as important for AI systems as it is for pharmaceuticals. AI systems, unlike traditional software, are dynamic, Maikel Leon, PhD, associate professor, practice in business technology, University of Miami, notes in a recent paper. Dynamic data dependencies and continuous learning mean that the outcomes an AI model predicted last month may not be those it predicts today for the same processes and conditions.
Additionally, there are tradeoffs between explainability and predictive performance, especially for highly complex models, Leon stresses. Therefore, biopharma manufacturers need to understand the hidden risks of the AI models they select and to develop effective interventions.
As Nakka, Reddy, and Gupta reiterate, “Sustainable good manufacturing practice deployment requires more than model accuracy: it requires data integrity, disciplined validation, and lifecycle governance.”
