A machine learning framework for polymeric nanoparticle design with zero-shot generalization across active pharmaceutical ingredients

The development of polymeric nanoparticles remains largely empirical, requiring extensive experimental screening to identify suitable formulation and processing conditions for each active pharmaceutical ingredient (API). Machine learning (ML) offers opportunities to support data-driven drug-delivery development, but existing approaches are often restricted to individual compounds and lack external validation. Here, we developed NANO-AI, a two-step ML framework integrating formulation variables, process parameters, and molecular descriptors to predict nanoparticle formation and key physicochemical attributes. The framework was trained on 925 experimental formulations encompassing nine physicochemically diverse APIs. In the first step, classification models predicted successful nanoparticle formation, achieving an area under the receiver operating characteristic curve of 0.823. In the second step, regression models predicted particle size, polydispersity index, zeta potential, and encapsulation efficiency, with R² values above 0.95 for size, zeta potential, and encapsulation efficiency. Interlaboratory experiments demonstrated robustness across operators and experimental environments. Importantly, zero-shot validation using previously unseen APIs showed that the framework could generate useful predictions across external compounds spanning different degrees of physicochemical similarity to the training APIs. These findings establish NANO-AI as a generalizable data-driven framework for rational polymeric nanocarrier development, potentially reducing empirical screening and supporting formulation development across chemically diverse therapeutic compounds.