Data-Driven Nanoparticle Design for Drug Delivery

Abstract

Drug delivery is a critical translational step in drug development that determines whether an active pharmaceutical ingredient can achieve adequate efficacy and safety in vivo and ultimately be translated into a drug product suitable for patient use. Despite significant advances in drug discovery, many promising compounds fail due to limitations in formulation, transport, and biodistribution rather than intrinsic biological activity. Nanoparticle-based drug delivery systems offer a means to engineer these processes by controlling drug stability, distribution, and interaction with biological barriers. However, nanoparticle design remains largely heuristic, relying on trial-and-error experimentation in high-dimensional and nonlinear formulation spaces.This dissertation develops and demonstrates a data-driven framework for nanoparticle design for drug delivery that integrates large-scale data curation, machine learning, laboratory automation, and experimental validation. The central hypothesis is that coupling curated data, learning-based models, and experimental validation within a unified pipeline enables more systematic exploration and optimization of nanoparticle formulations beyond intuition-driven approaches. First, we demonstrate that a high-fidelity database within a specific research niche can be constructed through systematic curation of the published literature. Using inorganic nanoparticles in preclinical cancer research as a representative case study, we assemble the most comprehensive dataset of its kind comprising studies with in vivo animal data. Retrospective statistical and machine learning analyses reveal historical nanoparticle design trends and quantitatively link formulation features to in vivo efficacy, highlighting underexplored regions of formulation space and motivating the need for standardized data reporting practices. Next, we show that laboratory automation provides a powerful and complementary approach for constructing high-quality in-house datasets to quantitatively explore formulation design space. Leveraging these data, we develop a bespoke hybrid kernel machine learning model tailored to the underlying data structure, enabling simultaneous materials selection and composition optimization in nanoparticle design. This approach outperforms conventional models in predicting nanoparticle formation in retrospective evaluations and is prospectively validated through the discovery and optimization of novel formulations with high drug loading and preserved biological performance. Furthermore, we extend the data-driven design framework to multicomponent nanoparticle systems using a permutation-invariant transformer architecture that respects the physical symmetries of mixture-based formulations. This model enables accurate prediction of nanoparticle formation and size, provides interpretable insight into intermolecular contributions, and supports uncertainty-aware design. Prospective experimental studies demonstrate the model’s ability to guide the discovery of functional multicomponent nanoparticles within the formulation regimes investigated. Finally, we explore the AI-guided design of targeted nanoparticles incorporating antibody components, representing a further step toward modeling biologically relevant interactions in nanoparticle drug delivery. The resulting AI-designed targeted nanoparticles exhibit enhanced receptor-specific cellular uptake and increased in vivo tumor accumulation relative to non-targeted formulations. Collectively, this dissertation establishes a data-driven pipeline for nanoparticle design that bridges computational prediction and experimental validation and is demonstrated across multiple nanoparticle platforms studied here. By emphasizing scalable data generation, model adaptation, and prospective biological evaluation, this work advances nanoparticle drug delivery toward a more systematic, reproducible, and translationally relevant design paradigm.