AI Formulation: Regulators Want Data

Pauric Bannigan, PhD, CSO and co-founder of Intrepid Labs, connected with PharmTech at AAPS PharmSci 360 2026 to discuss how computational tools are changing pharmaceutical formulation development.

Bannigan does not see mechanistic and data-driven models as competing approaches. Instead, he sees them converging in what are commonly called physics-informed machine learning models. A conventional model takes composition and tries to predict a single value, such as percent drug release. A physics-informed model takes in composition and the release experiment setup and fits a governing relationship such as the Noyes-Whitney equation. “There’s a lot of instances where this is a very, very successful strategy, and I think it aligns well with regulatory guidance as well,” he says.

Regulatory expectations, however, remain grounded in the laboratory. “In our experience, these models are useful decision support tools, but we’re not at the point yet, at least, where regulators are gonna take the prediction of a model as ground truth,” Bannigan explains. He says model outputs should serve as suggestions, which formulators still need to test and characterize experimentally.

The outlook is more encouraging for small and mid-sized companies without dedicated data science teams. Bannigan notes that over the past 6 to 12 months, the barrier to entry for machine learning on small data sets has dropped sharply. Chat-based software tools can now help formulation scientists build models from scratch. Together with free online resources and published literature, these tools offer a practical first step. That step builds on earlier industry efforts to apply AI and machine learning to solubility and bioavailability challenges.