Artificial intelligence can help anticipate the diminishing effectiveness of medications.

The polymorphism of molecular crystals may not seem like a topic suited for the kitchen, yet it can be relevant. For instance, chocolate that has been sitting on the shelf for too long appears suspicious: it doesn't seem spoiled, but something feels off. What has happened to its molecules?

The main component of chocolate is cocoa butter, and from a chemical standpoint, a bar that has been stored in the cupboard is still cocoa butter. However, its molecules have rearranged into a different, less desirable crystalline structure. But "polymorphic" chocolate isn't the worst scenario: it turns out that similar processes can weaken the effectiveness of medications.

Polymorphs refer to the various crystalline forms that the same substance can take depending on external conditions. For cocoa butter, for example, there are six such forms, and factories employ various tricks involving temperature changes to maximize the production of the fifth polymorph—this is the one that gives chocolate its shine, texture, and causes it to melt in your mouth while breaking with that characteristic chocolate sound. However, after prolonged storage, the fifth polymorph starts to transition into the less appetizing sixth form. What does this have to do with medications?

“In 1985, the substance rotigotine was discovered, and for a long time, only one polymorph of it was known. In 2007, the drug was approved in patch form for Parkinson's disease. A year later, it was found that there was another, more stable and less soluble polymorph, which caused a significant uproar and huge losses for the manufacturer: the drug was hastily withdrawn from the market for reformulation. Solubility is one of those properties that are crucial for the action of medications, but it depends not on the chemical composition but on the crystalline form that the active substance molecules take in a tablet or, as in this case, a patch,” explained Nikita Rybin, a researcher at the Skoltech Center for Artificial Intelligence.

Together with colleagues, Rybin published a study in the journal Physical Chemistry Chemical Physics research, supported by a grant from the Russian Science Foundation, proposing the use of so-called machine-learning potentials to accelerate the screening of substances for stable polymorphs, in order to prevent similar scandals with other medications in the future. The authors tested their approach on well-studied molecules of glycine and benzene, accurately predicting the known stable polymorphs of these substances while utilizing limited computational resources.

“Properties can be predicted directly through quantum-mechanical calculations. This is exactly what the winners of a recent competition, held annually by the Cambridge-based non-profit organization CCDC since the rotigotine incident, did,” Rybin explained. “However, this approach is impractical for pharmaceutical companies that need to screen millions of potential active substances. Quantum-mechanical modeling, as well as physical experiments, are applied at the final stage when the list of candidate substances has been narrowed down to at least a few dozen. Therefore, everyone is looking for ways to speed up the modeling process.”

One of the most promising approaches is machine-learning potentials for interatomic interactions. These are models trained on data from a small number of computational experiments performed on a smaller scale, yet with full quantum-mechanical accuracy. As a result, larger-scale models provide accuracy comparable to quantum-mechanical calculations, but with significantly simpler computations. If this intermediate step of machine learning is bypassed and calculations are done directly, the computations become unmanageable when reaching a scale where the physical properties of interest are visible.

The research team, co-authored by Professor Alexander Shapeev from the Skoltech Center for AI and head of the Laboratory of Artificial Intelligence Methods for Materials Development, has already applied machine-learning potentials in the search for materials for nuclear energy and aerospace. Now, shifting from inorganic to molecular crystals, the team has demonstrated that these potentials can also be beneficial in drug development, accelerating the screening of polymorphic forms of active substances by a thousandfold or more.

A thorough examination of the physical properties of active substances in the form of tablets or patches will allow manufacturers to anticipate issues related to insufficient solubility, degradation due to heating, exposure to air, and more, thus avoiding unpleasant surprises. To achieve this, the Skoltech research group plans to move on to more complex and pharmacologically significant compounds and refine the methodology to account for humidity and other environmental parameters.