Data for Data-driven discovery of potent small molecule ice recrystallisation inhibitors

Current programmes to identify materials capable of modulating ice growth are guided solely by iterative changes and human intuition, with a major focus on macromolecules (proteins or polymers). This process is fundamentally constrained by a poor understanding of the mechanisms and underlying structure-activity relationships.

Here, we overcome this barrier by constructing machine learning models capable of predicting the ice growth inhibition activity of small molecules. Due to current limitations in experimental throughput, we leverage ensemble models which combine state-of-the-art descriptors with domain-specific features derived from molecular simulations. When applied to virtually screen a commercial compound library, these models successfully predicted novel ice recrystallisation inhibitors that are experimentally verified to function at low millimolar concentrations.