Representing Molecules in Their Physical Context through Molecular Dynamics

Predictive machine learning models in chemistry predominantly treat molecules as isolated entities. While molecular graphs provide a powerful inductive bias by encoding atomic identity and intramolecular connectivity, they lack an equivalent way to represent collective intermolecular interactions that govern many condensed-phase phenomena. For example, the local arrangement of molecules in a liquid mixture can promote hydrogen-bond networks that strongly influence macroscopic properties, but conventional models can only infer such effects indirectly from property labels. Here, we show that molecular dynamics (MD) can serve as a physics-grounded virtual environment for observing and encoding this molecular context. We introduce atom-in-context (AIC) representations, which embed molecules within dynamically sampled condensed-phase neighborhoods through explicit intermolecular graph connectivity. Using dynamic viscosity of liquid mixtures as a representative property, AIC performs comparably to isolated representations for approximately linear mixtures, but substantially improves generalization to unseen strongly non-ideal mixtures, increasing R² from 0.56 $\pm$ 0.24 to 0.82 $\pm$ 0.05 and reducing MAE from 0.42$\pm$0.13 to 0.24$\pm$0.02 log cP. Learned representations of the same molecule vary across mixture environments, demonstrating that MD can provide context-dependent molecular representations that encode not only what molecules are, but the environments in which they exist, opening a way to encode molecular systems.