Graph Learning on Ensembles of Cyclic Peptides: An Investigation of Molecular Ensemble Modeling
Teaching AI to predict how molecules behave by modeling their shape-shifting in solution
Molecules like cyclic peptides constantly twist and shift between different shapes in liquid, yet current AI models predict their properties from just a single frozen structure. Researchers built EnsembleEGNN, a neural network that encodes multiple conformations of the same molecule simultaneously, then pools them into a single prediction. The model outperformed sequence-only baselines, reaching 74% correlation with experimental properties when trained end-to-end—something that completely failed without access to 3D structural ensembles.
Drug discovery relies on predicting how molecules will behave, and current shortcuts miss a crucial reality: molecules are not rigid shapes but dancing ensembles. This approach could speed up screening for new peptide drugs by accurately predicting their properties from realistic representations of how they actually move in cells. It opens a path toward foundation models that capture molecular dynamics, moving beyond static snapshots that have limited predictive power.