Digital Materials is launching a new research direction focused on developing computational solutions for the pharmaceutical industry. Our work focuses on molecular crystal modeling, with the goal of creating tools that can accelerate the development and screening of pharmaceutical compounds.

One of the first developments in this area addresses the problem of identifying stable molecular crystal forms. The methodology is designed to accelerate the search for stable polymorphs. Polymorphism is the ability of the same substance to form different crystal structures depending on external conditions.

This phenomenon is particularly important in pharmaceuticals because different crystal forms of the same compound can have different solubility, bioavailability, and stability. These properties can directly affect the efficacy and safety of a drug.

“In 1985, rotigotine was discovered, and for many years only one of its polymorphs was known. In 2007, the drug was approved as a transdermal patch for Parkinson’s disease. One year later, however, another polymorph was discovered that was more stable and less soluble. This caused major problems and significant financial losses for the manufacturer: the product had to be urgently recalled and reformulated.

Solubility is one of the properties that is critical to drug performance but depends not on chemical composition alone, but also on the crystal form adopted by the molecules of the active ingredient in a tablet or, in this case, a patch,” explains Nikita Rybin, CEO of Digital Materials.

The results of this research were published in Physical Chemistry Chemical Physics. The developed methodology can help pharmaceutical companies improve both the speed and reliability of drug development and reduce the risk of problems similar to those encountered with rotigotine. The approach has already been successfully applied to predict stable polymorphs of glycine and benzene, demonstrating its effectiveness.

“Material properties can be predicted directly using quantum-mechanical calculations. This was essentially the approach taken by the winners of a recent competition organized by the Cambridge Crystallographic Data Centre, which has been running such blind tests regularly for many years,” says Nikita Rybin. “However, this approach is not suitable for pharmaceutical companies that need to screen millions of potential active compounds.

Quantum-mechanical modeling, just like physical experiments, is typically introduced only at the final stage, once the number of candidate compounds has already been narrowed down to a few dozen. This is why there is such strong demand for faster computational methods.”

The proposed methodology is based on MTP (Moment Tensor Potential) machine-learned interatomic potentials. These models are trained on a limited number of smaller-scale calculations performed with quantum-mechanical accuracy. Once trained, they can be applied to much larger systems while achieving accuracy comparable to quantum-mechanical calculations at a substantially lower computational cost.

Direct large-scale quantum-mechanical calculations are extremely computationally demanding. Using machine-learned interatomic potentials can therefore significantly accelerate drug development by making the screening of polymorphic forms of active pharmaceutical ingredients thousands of times more efficient.

Careful computational assessment of the physical properties of active ingredients in tablets or patches can help pharmaceutical companies identify potential problems in advance, including insufficient solubility or degradation of drug quality under exposure to temperature or air. Future work will focus on more complex, pharmacologically relevant compounds and on extending the methodology to account for factors such as humidity and other environmental conditions.

At the same time, we continue to improve computational methods for molecular crystals. In particular, we have extended MTP machine-learned interatomic potentials by explicitly incorporating van der Waals and Coulomb interactions, both of which are particularly important for organic compounds. The results were published in The Journal of Chemical Physics.

In parallel, we investigated optimized approaches based on van der Waals density functionals, including DF1 and DF2, for accurate modeling of molecular crystal structures. Together, these studies provide a foundation for a new level of accuracy and efficiency in computational molecular crystal research.