Molecular crystal polymorphism does not sound like something you want in your kitchen, but you’ve probably had experience with it. Like that time you left a chocolate bar lying around for months, only to find it looking weird afterward. Has someone been messing with its molecules? Chemically speaking, it’s still chocolate — that is, mostly cocoa butter — but the molecules have gradually rearranged themselves into a different crystal structure while you weren’t looking. And as if chocolate wasn’t bad enough, this can actually happen to medications, causing them to lose efficacy.

Polymorphs are distinct molecular crystal structures formed by one and the same compound under different circumstances. Cocoa butter, for example, comes in six such forms, and chocolate makers employ sophisticated temperature manipulations to maximize the “tasty” polymorph content in your chocolate bar, making sure it is shiny, smooth, snappy, and melts in your mouth. Even so, prolonged storage, particularly at suboptimal conditions, causes that polymorph to “mutate” into its less palatable counterpart. All of this applies to the molecules in pills, too.

“Since 1985, drug manufacturers had only been aware of one polymorphic form of rotigotine — a medication prescribed for the treatment of Parkinson’s disease. In 2008, however, the discovery of a significantly more stable and less soluble polymorph prompted a massive drug recall with huge economic losses and public health repercussions. Solubility is one of those properties that are essential for the medication to have its intended effect, and yet it depends on the crystal structure assumed by the molecules in the pill or, in this case, the transdermal patch, rather than the drug’s chemical makeup,” Research Scientist Nikita Rybin from Skoltech AI commented.

Together with his colleagues, Rybin published a study in Physical Chemistry Chemical Physics, backed by Russian Science Foundation Grant No. 23-13-00332, proposing the use of so-called machine-learned interatomic potentials to accelerate polymorph screening and avoid similar debacles with other pharmaceuticals in the future. Using the well-studied glycine and benzene molecules to test the technique, the team correctly predicted the stable polymorphs of these two compounds using fairly modest computational resources.

“You can predict it the hard way, by doing direct quantum mechanical computations. Indeed, such a brute force approach has recently triumphed at the Crystal Structure Prediction Blind Test contest, held by the Cambridge-based nonprofit CCDC every year since the infamous rotigotine story,” Rybin said. “This is not feasible for pharmaceutical companies, though. They have to screen millions of drug candidates, and full quantum mechanical simulations — just like actual wet experiments — are only an option for, perhaps, dozens of preselected molecules. So, people are exploring ways to speed up this procedure.”

Among them are machine-learned interatomic potentials. These are trained on the output of smaller-scale models formulated with quantum mechanical accuracy and subsequently used to sidestep overwhelmingly difficult calculations. If it weren’t for machine learning, the direct fundamental calculations would have gotten way too demanding computationally once the researchers went from the manageable small-scale model to a scale large enough for the relevant physical properties to emerge.

Headed by study co-author Professor Alexander Shapeev, the Laboratory of Artificial Intelligence for Materials Design at Skoltech AI recently deployed machine-learned potentials to speed up the search for salts for next-generation nuclear power plants, as well as industrial metal alloys for aerospace tech. Extending its domain of application from inorganic to molecular crystals, this time the team showed the technique could be harnessed for drug design, too, accelerating molecular crystal polymorph screening by a factor of 1,000 or more.

By thoroughly testing the physical properties of the active compounds of drugs in the form of pills or patches, medical research centers and R&D departments of pharma companies will be able to check for insolubility issues, potential degradation in open-air conditions or upon heating, etc., and avoid possible mishaps, such as the one that involved rotigotine. To make this a reality, the Skoltech team intends to move on to more intricately structured pharmaceutically significant molecules and to develop the proposed technique so as to account for ambient humidity and other environmental parameters.