In a proof-of-concept study, the team used the platform to design nanoparticles that successfully delivered venetoclax, a chemotherapy drug known for being difficult to encapsulate. They also optimized the formulation of a second anti-cancer nanoparticle, demonstrating the platform’s versatility.
AI has significantly accelerated early-stage drug discovery, helping researchers predict the biological, chemical, and physical properties of potential therapeutic molecules. Many of these AI-generated candidates are already advancing through clinical trials. But when it comes to the next phase (formulation and delivery), AI's role remains relatively underdeveloped.
When you’re creating a nanoparticle, how well it works doesn’t just depend on the recipe, but also on the quantity of the various ingredients, including both the active drug and inactive materials. Existing AI platforms can only handle one or the other, which limits their overall effectiveness.
Zilu Zhang, Ph.D. Student, Biomedical Engineering, Duke University
That’s where the new platform comes in. According to Zhang, most existing AI tools can either recommend what materials to use or how much of each to include—but not both. That limitation often leads to less-than-optimal nanoparticle designs.
AI can help us identify promising delivery molecules, but if you don’t mix them with the drug at a certain ratio, they won’t form a stable nanoparticle. If we can identify the optimal mixture ratios, then we can form the particles and maintain their stability.
Daniel Reker, Assistant Professor, Biomedical Engineering, Duke University
Beyond their limited ability to account for both ingredients and their proportions, existing approaches come with additional hurdles. More advanced AI models can identify material properties and effective ratios, but they typically require enormous datasets to function well. On the other hand, simpler models can operate with less data but often struggle to distinguish between similar compounds.
Reker and Zhang aim to overcome these limitations with their new AI-guided platform, TuNa-AI (Tunable Nanoparticle platform).
To train it, they used an automated liquid-handling system to generate a dataset of 1275 unique nanoparticle formulations. Each formulation combined various therapeutic molecules with excipients—non-active substances like preservatives, stabilizers, or coloring agents that enhance a drug’s physical properties and absorption.
By using robotics, we were able to combine many different ingredients in many different recipes very systematically. Our AI model was then able to look at that data for how different materials perform under different conditions and extrapolate that knowledge to select an optimized nanoparticle.
Zilu Zhang, Ph.D. Student, Biomedical Engineering, Duke University
The team found that TuNa-AI led to a 42.9 % increase in successful nanoparticle formation compared to conventional methods. As a proof of concept, they used the platform to formulate a nanoparticle that more effectively encapsulated venetoclax, a chemotherapy drug used to treat leukemia. The resulting nanoparticles showed improved solubility and were significantly more effective at halting leukemia cell growth in laboratory tests than the non-encapsulated drug.
In a second case study, the AI-driven platform helped reformulate another chemotherapy drug by reducing the use of a potentially carcinogenic excipient by 75 %, all while maintaining the drug’s efficacy and improving its biodistribution in mouse models.
“We showed that TuNa-AI can be used not only to identify new nanoparticles but also optimize existing materials to make them safer,” said Zhang.
In addition to expanding their platform to work with a broader range of biomaterials for therapeutic and diagnostic applications, the team is actively collaborating with researchers and clinicians, both at Duke and beyond, to apply the technology to drug delivery challenges in hard-to-treat diseases.
“This platform is a big foundational step for designing and optimizing nanoparticles for therapeutic applications. Now, we’re excited to look ahead and treat diseases by making existing and new therapies more effective and safer,” said Reker.
Journal Reference:
Zhang, Z., et al. (2025) TuNa-AI: A Hybrid Kernel Machine To Design Tunable Nanoparticles for Drug Delivery. ACS Nano. doi.org/10.1021/acsnano.5c09066