AI learns to design a drug’s release timetable

A long-acting medicine needs a timetable as well as a dose. Packing a drug into a microscopic polymer sphere creates a structure through which water enters and drug escapes. The manufacturing process helps determine how that structure behaves. An early rush of release can be followed by a very different rate later on.[3]

Ying Qin and colleagues described E-MAF in Nature Communications on 15 September. It searches formulation and manufacturing choices for a desired release pattern.[1]

The process becomes part of the recipe

The physical connection predates AI. In a 2005 experimental study, researchers used nuclear magnetic resonance and confocal microscopy to examine drug-loaded microspheres made from poly(lactic-co-glycolic acid), or PLGA, immersed in water. Changing how the particles were dried altered their permeability to water; changing the temperature during preparation altered how the internal pore structure evolved. The finished particle carries a history of how it was made.[3]

That is why a list of ingredients is incomplete. Two preparations can aim at the same nominal composition while producing different pathways for material to move through. For a design system, the manufacturing choices belong among the inputs used to predict a release curve.[3]

The released E-MAF program makes the search concrete. A candidate contains numerical formulation descriptors together with choices such as encapsulation and solidification method. A learned predictor estimates how much drug will be retained. Another estimates the fraction released at specified times. These predictions let the program compare many possible preparations before selecting ones to make.[2]

Searching backwards from a desired curve

The calculation runs in two directions. The predictor goes forwards, from a proposed preparation to its expected behaviour. The optimiser goes backwards, starting with desired release values and searching for preparations whose predictions resemble them. In the released demonstration, the target consists of nine time points read from a design spreadsheet.[2]

The program first screens candidates for predicted drug retention. It then uses a genetic algorithm: candidate preparations are recombined and randomly varied, and those with a closer predicted fit are favoured. The allowed parameter ranges restrict where it searches. Its fitness function is based on mean absolute error across the target curve, so smaller average discrepancies receive better scores.[2]

This arrangement lets a scientist ask for a shape: little release initially, a sustained middle phase, then a taper. A single endpoint would be a weaker instruction. Two preparations might release the same total amount by the final day while exposing the body to very different amounts along the way.

What changed in the laboratory

For goserelin, the model’s interpretation suggested halving internal drug concentration from 125 to 62.5 mg/mL. Other settings stayed fixed. Two-hour release fell from 3.39 ± 1.04% to 1.55 ± 0.13%, with comparable particle size and appearance. Means and standard deviations came from three independent laboratory release measurements. Lower internal concentration reduces the outward concentration difference that drives diffusion.[1]

Using those reported means, our calculation gives a 1.84-percentage-point reduction, equivalent to 54.3% less early release. The controlled change tests a specific formulation adjustment.[1]

The scientific value of such a test is the link between a prediction and something a laboratory can change. A ranked feature alone would leave the explanation at association. Making the altered preparation and measuring its release checks whether the proposed adjustment produces the intended direction of change.

A separate AI-designed risperidone preparation matched its target curve with a 5.13-percentage-point mean error under accelerated testing at 45°C. In rats, it and the commercial comparator Rykindo both sustained serum drug concentrations for about a month, with six animals per group. These are formulation and animal results, without demonstrated human clinical benefit.[1]

A curve is a design instruction

The software also calculates discrepancies separately for early, middle and late portions of its nine-point target. That makes it possible to inspect where a candidate misses the desired behaviour, even when its overall average looks good. The released search retains multiple candidates and ranks parameter variation across search settings, giving researchers a set of preparations to inspect.[2]

For the scientist, the useful output is a preparation that can be made and a schedule that can be measured. Each time point asks a concrete question: how much has left the particle by now? Treating that sequence as the design target connects AI’s search to the physical job of a controlled-release medicine.

Report an error or suggest a correction