AI Predicts Drug Release With Up to 40% Fewer Errors, Cutting Nearly a Day of Lab Testing

A new medication can spend years in the lab before a single patient ever takes it. For controlled-release medicines, the kind engineered to let a drug out slowly over hours or days, one early bottleneck has nothing to do with whether the drug works and everything to do with timing: figuring out how fast the drug leaks out of the patch, capsule, or film built to release it. That single measurement can eat up days of round-the-clock lab work, long before anyone weighs a clinical trial. Researchers at Brown University have now shown a way to predict how the drug behaves across the rest of that release window, a stretch that ran 48 hours in the lab tests, from a sliver of the usual data, trimming up to nearly a full day off the process for some materials.

Their tool leans on a kind of artificial intelligence called a Physics-Informed Neural Network, or PINN, which blends machine learning with the actual laws of physics. Rather than feeding the software mountains of measurements and hoping it spots a pattern, the team taught it the scientific rule that governs how substances spread through a material. Anchored by that rule, the AI could train on a small batch of early readings and still call the drug’s later behavior correctly. Across the materials tested, it cut average error by as much as 40 percent compared with the math models labs have relied on for decades.

Published in the Journal of Drug Delivery Science and Technology, the work points to a hybrid approach that could reshape how drug companies test these controlled-release medicines.

How AI Predicts Drug Release From Less Data

Standard drug release models have anchored this field for decades. They use math formulas to describe how a drug spreads out of its carrier material, a bit like a drop of food coloring easing through a glass of water, only inside an engineered thin film or capsule.

Those formulas are handy, but they carry a big assumption: perfectly smooth, uniform conditions. Actual delivery systems are messier. A film can be flat, wrinkled, or crumpled in ways that change how a drug moves through it, and once the geometry gets complicated, the old formulas start to slip.

PINNs get around this by training a neural network, the engine behind many familiar AI tools, with one important change. Instead of letting the software work everything out from data alone, the researchers baked the law of how substances spread straight into its training. Every guess the model makes gets checked against that rule, which keeps it tethered to reality even when it has very little data to go on.

Three Film Shapes, One Model

To put the method through its paces, the team borrowed data from an earlier published study that tracked how a test compound seeped out of three kinds of ultrathin graphene oxide films: flat, wrinkled in one direction, and crumpled in two. Each shape releases the compound on its own schedule, since geometry changes the escape routes.

That dataset held 15 measured time points for each film. Researchers trained the AI on shrinking slices of it, from as many as 14 points down to just 2, to find out how little it could get by on and still forecast the rest of the release curve accurately.

For the flat film, the PINN reached solid accuracy with 9 data points, about 120 minutes of release readings. Older math models needed 12 to 13 points, or 1 to 1.5 days of lab time, to match it. For the wrinkled and crumpled films, the PINN hit the same mark at 11 points, roughly 12 hours in, while the classic models again wanted 12 to 13. Depending on the film, that adds up to 12 to 36 hours of testing saved.

Researchers also ran the process in reverse, using the AI to estimate how easily molecules travel through each film. That backward step relied on a simplified one-dimensional model, so its numbers are best read as rough, model-based estimates rather than an exact map of how molecules move through the films’ shapes.

Handling Messy Lab Data

Real lab readings are never spotless. Temperature drift, instrument limits, and plain human handling all nudge the numbers around. To mimic that, researchers deliberately spiked their data with simulated noise and watched how each model coped.

Standard PINNs wobbled a bit under noise, with wider swings in their predictions. So the team built a sturdier version, a Bayesian PINN, using a technique called Monte Carlo Dropout that produces a range of answers along with a confidence estimate for each one. Under noisy conditions it logged lower error and tighter, steadier confidence bands than an ensemble of 50 ordinary PINNs. The tradeoff, which the authors flag plainly, is that the Bayesian version costs more computing power because of the extra sampling it runs.

A Faster Path for Drug Release Research

Drug development is slow and pricey, and early testing of delivery systems is one of the choke points along the way. Anything that shortens the lab clock without giving up accuracy earns a look. The authors say their framework may carry over to delivery systems beyond the three films tested here, though they stop short of claiming it has been proven there yet.

Forecasting the rest of a drug’s release curve from only its first few readings is a real shortcut, not a rounding error. Should the method hold up under wider testing, it could speed the trip from a promising compound to the patients waiting on it.

Disclaimer: This article summarizes a peer-reviewed research paper for a general audience and is based on a pre-publication version; the final published text may differ. It is intended for informational purposes only and is not medical, pharmaceutical, or investment advice. The AI method described was tested on a single previously published dataset of graphene oxide films and has not yet been validated across the broader range of drugs, materials, and conditions used in real-world drug development. Findings should be read as early-stage and in need of wider testing before any clinical or commercial application.

---