Discovering interpretable drug formulation behavior patterns via a mechanistic-augmented conditional variational autoencoder.

Formulation composition, processing conditions, and their combined effect on drug solubility and particle size are not fully understood. This is due to a complicated network of interactions dependent on the conditions, which are usually investigated by trial-and-error testing. A question that has been left practically unanswered is whether the experimental formulation data can still find patterns of structured and interpretable behavior that are beyond mere prediction. This paper tries to answer whether the latent generative modeling can structurize formulation knowledge in a continuous, regime-aware manner. 114 niosome formulation samples were mined systematically from 17 publications based on the PRISMA framework. Inputs to model the encapsulated drug efficiency and particle size were 11 drug, formulation, and processing variables. To develop a structured latent understanding of formulation behavior, a mechanistic-augmented conditional variational autoencoder was used. The discovered latent space became a continuum with regimes overlapping, smooth changes, and feature, response relationships being dependent on the context. In a sense, formulation behavior can be considered as a structured latent landscape that can be used for regime-aware analysis and the generation of hypotheses in data-driven formulation research.