FlexENN: A Graph Neural Network for Binding Energy Prediction of Globular and Intrinsically Disordered Proteins

Intrinsically disordered proteins (IDPs) drive a large fraction of cellular signaling, transcription, and assembly through interfaces that often lack a defined geometry, making the prediction of their binding energies challenging for methods developed primarily for folded complexes. Here, we introduce FlexENN, an interface-focused message-passing graph neural network framework comprising two independently trained models: FlexENN-F for folded protein complexes and FlexENN-I for complexes containing a disordered partner. The FlexENN architecture constructs a local interface graph in which each residue node integrates the relative geometry, flexibility-related structural descriptors, and sequence information, allowing the network to learn structure-affinity relationships from a single resolved bound conformation. We also construct an IDP-containing dataset for training the FlexENN-I model. Benchmarked across folded complexes and IDP-containing systems, including tubulin-tubulin interfaces within microtubules (MTs) and MTs in complex with microtubule-associated proteins (MAPs), with reference binding affinity values extracted from our experiments, the models reproduce physically reasonable energetic profiles, including the stronger interdimer relative to the lateral MT interactions, while feature-ablation analyses reveal system-dependent contributions of selected descriptors to the predictions. More broadly, this work delivers an accurate, structure-aware approach to predicting binding free energies for complexes of folded protein and complexes that contain ordered or partially ordered IDP regions for which a sufficiently well-defined bound structure is available.