Graph-based intermolecular interaction prediction enables rational design of co-crystals

Abstract

Co-crystal engineering has emerged as a promising strategy to tailor the properties of organic materials, yet the mechanistic principles governing co-crystallization remain underexplored. Here, we systematically analyze co-crystal formation using intermolecular interaction as the criterion from a thermodynamic perspective. Building on these energetic insights, a computational framework integrating intermolecular interaction analysis and machine learning methods to enable rational design of co-crystals was established. First, anomaly detection via local outlier factor (LOF) algorithms preliminarily screened out abnormally designed co-crystals based on energy decomposition analysis of existing co-crystals. Next, a comparative energy benchmark was established, requiring that the intermolecular interaction between the two designed molecules is more stable than that in their individual component, ensuring thermodynamic feasibility. Furthermore, a graph neural network (GNN) model was implemented to rapidly predict intermolecular interaction profiles, enabling efficient co-crystal design. This machine learning-enhanced workflow accelerated the identification of two novel urea-based co-crystals, which were experimentally synthesized and characterized. Our findings not only quantify the energy steering supramolecular assembly but also break through the traditional method of co-crystal classification, providing new perspectives for the mechanism and prediction of co-crystal formation.