In drug delivery, "AI" covers two quite different things: machine learning models that predict how a carrier such as a nanoparticle will behave before it is made, and automated dosing systems that adjust a drug's delivery from sensor readings. The first is mostly at the research stage and depends on small, hard-to-compare datasets. The second has randomized trial evidence for at least one disease, with limits on what that evidence covers.
Key takeaways
- A 2026 peer-reviewed review describes machine learning that predicts nanoparticle size, drug loading, release, and stability, and names data scarcity and lack of standardization as main obstacles.
- A 2019 randomized trial of 168 people with type 1 diabetes reported 71 percent time in the glucose target range with a closed-loop system, against 59 percent with a sensor-augmented pump.
- The FDA's January 2025 draft guidance on AI in drug development asks sponsors to show that a model is credible for its specific context of use.
- Predictions from a model are not evidence of safety or effectiveness in people; that requires laboratory, animal, and human studies.
Designing carriers with machine learning
Many drugs reach their target through a carrier, such as a lipid nanoparticle, liposome, or polymer particle. Choosing materials and process settings by trial and error is slow. A review in Acta Pharmaceutica Sinica B (volume 16, February 2026) surveys how machine learning is applied to this problem. The authors note that more than 90 nanomedicines are approved and that formulation development remains complex.
The review describes models that predict particle size, polydispersity, and zeta potential, with neural networks the most common choice. Others estimate encapsulation efficiency and drug loading, forecast release profiles using tree-based methods, classify stability, screen ionizable lipids, and read microscopy images. One cited approach uses a digital twin of a nanosuspension milling process to tune manufacturing settings.
Why the models are hard to trust yet
The same review lists problems that limit how far the predictions can be taken. Data are limited, fragmented, and of several types, which can lead to overfitting or underfitting and weak generalization. Characterization methods differ, for example dynamic light scattering versus electron microscopy, so results from separate laboratories are hard to compare. Many models are black boxes, which complicates troubleshooting and regulatory acceptance. Training data may also be unrepresentative, and the authors give the example of mostly male subjects.
Above all, the review stresses that predictions must be validated in vivo and in human studies, and that scale-up, quality control, and regulatory requirements remain hurdles. The authors propose open nanomaterial databases, explainable models, federated learning, and transfer and active learning for small datasets. These are proposals, not established practice.
Automated dosing: one trial and its limits
The second meaning concerns systems that adjust delivery of a drug automatically. The best-studied example is closed-loop insulin delivery. In a trial published in the New England Journal of Medicine on October 31, 2019, 168 patients with type 1 diabetes aged 14 to 71 were assigned 2 to 1 to a closed-loop system or a sensor-augmented pump. Time in the 70 to 180 mg per deciliter (3.9 to 10.0 mmol per liter) glucose range rose from 61 percent at baseline to 71 percent in the closed-loop group, and stayed at 59 percent in the control group. The adjusted difference was 11 percentage points. No serious hypoglycemic events occurred in either group, and one episode of diabetic ketoacidosis occurred in the closed-loop group.
The trial lasted six months, compared the system with another device, not with no technology, and measured glucose outcomes. The abstract describes closed-loop control, not machine learning, so whether such a system counts as AI depends on the definition used. These results do not apply to other drugs or to other populations, and dosing decisions belong with a patient's care team.
What regulators ask of AI models
The U.S. Food and Drug Administration issued a draft guidance in January 2025 (docket FDA-2024-D-4689) on using AI to support regulatory decisions about the safety, effectiveness, or quality of drugs. The FDA page labels it nonbinding and not for implementation. It proposes a seven-step, risk-based credibility framework. Credibility is judged for a defined context of use, which is the specific role and scope of the model.
| Step | What the sponsor does |
|---|---|
| 1 | Define the question the model addresses |
| 2 | Define the model's context of use |
| 3 | Assess the risk of the model |
| 4 | Plan how to establish credibility within that context |
| 5 | Carry out the plan |
| 6 | Document results and deviations |
| 7 | Decide whether the model is adequate for the context of use |
The draft does not address AI used in drug discovery, or for operational efficiencies that do not affect patient safety, drug quality, or the reliability of study results. A formulation model that only shortlists candidates may therefore fall outside it, while a model whose output supports a quality decision may fall inside, and the FDA encourages sponsors to ask early when unsure.
Reading claims about AI drug delivery
- Is the claim about a prediction made in a computer, a laboratory result, an animal study, or a human trial?
- How large and varied was the training dataset, and was the model tested on data it had not seen?
- Does the term "AI" describe machine learning, or a fixed control algorithm?
The research and regulatory documents summarized here are background reading. Nothing in them or in this text is a recommendation of any drug, device, or dose, and none of it replaces professional care. Decisions about medicines belong with a patient and their clinician or pharmacist.
FAQ
Is there an AI-designed drug delivery system on the market?
No such product could be verified for this article. The review describes prediction and screening methods in research, and it says translation to approved products still needs validation and regulatory steps.
Does closed-loop insulin delivery use AI?
The trial abstract describes closed-loop control that automates insulin delivery and does not call it machine learning. Whether it counts as AI depends on how the term is defined.
Can a model replace laboratory testing of a new formulation?
The evidence reviewed does not support that. Authors of the 2026 review say predictions need in vivo and human validation, and that data limits weaken generalization.
Is the FDA's AI guidance for drugs final?
The FDA page checked in October 2026 labels it draft guidance with nonbinding recommendations. Readers can check that page for any later change.
Related reading
- Health AI rules in Canada
- AI in scientific research
- Training datasets for AI
- How neural networks learn
More in AI in health.
Last verified: October 2026
Sources
- Shen C, et al. Machine learning empowered formulation design, optimization and characterization of nanoparticulate drug delivery systems: current applications, challenges, and future perspectives. Acta Pharmaceutica Sinica B 16(2), 2026
- Brown SA, et al. Six-month randomized, multicenter trial of closed-loop control in type 1 diabetes. New England Journal of Medicine 381(18), 2019
- Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products (draft guidance), U.S. Food and Drug Administration, January 2025