AI predicts drug release with up to 40% fewer errors, saving lab testing by nearly a day

A new drug can take years in the lab before a single patient takes it. For controlled-release drugs, which are designed to slowly release the drug over hours or days, one of the early bottlenecks has nothing to do with whether the drug works, and everything to do with timing. It’s about understanding how quickly drugs leak out of the patches, capsules, or films that are made to release them. That single measurement can consume days of around-the-clock laboratory work, long before it can be evaluated in a clinical trial. Researchers at Brown University now show how they can use some of their regular data to predict how a drug will behave over the remainder of its release period (48 hours in lab testing), cutting the process by nearly an entire day for some materials.

Their tools are based on a type of artificial intelligence called Physics-Informed Neural Networks (PINN), which blends machine learning with real physics. Rather than feeding the software a bunch of measurements and hoping it discovers patterns, the researchers taught it the scientific laws that govern how substances diffuse through materials. Based on this rule, the AI ​​can be trained on small batches of initial readings and correctly recall subsequent drug behavior. Across the materials tested, we reduced the average error by as much as 40% compared to the mathematical models that labs have relied on for decades.

Published in Drug Delivery Science and Technology Journalthis study demonstrates a hybrid approach that has the potential to reshape how pharmaceutical companies test these controlled-release drugs.

How AI predicts drug release from less data

Standard drug release models have supported the field for decades. They use mathematical formulas to explain how the drug diffuses out of the carrier material. This is similar to a drop of food coloring slowly passing through a glass of water, only inside an artificially created thin film or capsule.

Although these formulas are useful, they make a major assumption: perfectly smooth and uniform conditions. The actual delivery system is more complex. The film can be flat, wrinkled, or crumpled, changing the way the drug passes through it. As shapes become more complex, the old formula begins to break down.

PINN gets around this problem by training neural networks, the engine behind many familiar AI tools, but with one important change. Rather than letting the software calculate everything from data alone, the researchers incorporated the laws of how matter spreads directly into the training. Every inference the model makes is checked against its rules, so the model is tied to reality even when there is little data.

3 film shapes, 1 model

To put this method into practice, the researchers borrowed data from a previously published study that tracked how a test compound oozed out of three types of ultrathin graphene oxide films: flat, wrinkled in one direction, and wrinkled in two directions. Each shape releases compound on its own schedule, as the geometry changes the escape route.

That dataset included 15 measurement time points for each movie. The researchers trained the AI ​​to shrink the slice from a maximum of 14 points to just 2 points, to see how much it would take, and accurately predict the rest of the release curve.

For flat film, PINN reached solid accuracy with 9 data points and approximately 120 minutes of release reading. Matching the old mathematical model required 12 to 13 points, or 1 to 1.5 days of experimental time. For wrinkled and wrinkled films, PINN reached the same mark at 11 points in about 12 hours, while the classic model again required 12-13 points. Depending on the film, you can save a total of 12 to 36 hours of testing.

The researchers also used AI to reverse the process and estimate the speed at which the molecules passed through each film. Because this regression step relies on a simplified one-dimensional model, its numbers are best read as rough model-based estimates rather than precise maps of how molecules move within the membrane geometry.

Handling messy lab data

Actual laboratory measurements are never perfect. Temperature drift, equipment limitations, and simple human handling all adjust the numbers. To mimic that, the researchers intentionally added simulated noise to the data and observed how each model coped.

The standard PINN wobbled a bit under the noise and the predictions were more erratic. So the team built a more robust version, Bayesian PINN, using a technique called Monte Carlo dropout that generates different answers and confidence estimates for each. Under noisy conditions, lower errors and tighter and more stable confidence bands were recorded than for an ensemble of 50 regular PINNs. The trade-off that the authors explicitly point out is that the Bayesian version performs additional sampling and therefore requires more computational power.

A faster path to drug release research

Developing drugs is slow and expensive, and early testing of delivery systems is one of the challenges in the process. Anything that shortens laboratory clocks without sacrificing accuracy gets attention. The authors say their framework may transfer to distribution systems other than the three movies tested here, but they do not claim that it has been proven there yet.

Predicting the rest of a drug’s release curve from only the first few measurements is a real shortcut, not a rounding error. If the method survives extensive testing, it could speed the passage of promising compounds to waiting patients.

Disclaimer: This article is a general audience summary of a peer-reviewed research article and is based on a pre-publication version. Final published text may vary. This is for informational purposes only and does not constitute medical, pharmaceutical, or investment advice. The described AI method was tested on a single previously published dataset of graphene oxide films and has yet to be validated across the wide range of drugs, materials, and conditions used in real-world drug development. The findings should be interpreted as early stage and require extensive testing before clinical or commercial application.

---