A Comprehensive Atlas and Machine-Learning Framework for Predicting IDR-Protein Binding Affinity

Intrinsically disordered regions (IDRs) enable pervasive, regulatable protein interactions, yet predicting their binding affinities remains challenging because disorder permits heterogeneous interfaces and context-dependent recognition. Here we introduce IBPC-Kd, a curated dataset of 1,785 IDR-ordered protein complexes with experimentally measured dissociation constants (Kd), spanning more than six orders of magnitude in affinity and covering diverse IDR lengths and partner classes. Across this resource, we find that global interfacial shape complementarity is a consistent correlate of affinity, followed by the structural order of the partner surface and a systematic electrostatic asymmetry between IDRs and their binders. Local physicochemical patterning modulates binding in a partner-dependent manner. Guided by these observations, we develop IDRBindNet, a graph-transformer model that integrates protein language model embeddings with interface geometry and residue-level chemical context to predict dissociation constants from IDR-partner complexes. IDRBindNet achieves state-of-the-art accuracy on held-out complexes (R2 up to 0.911) while implicitly learning biologically relevant attention patterns corresponding to partner disorder and geometric fit. Notably, the framework generalizes through external validation on a set of recently reported de novo designed binders targeting the human surfaceome, showing encouraging concordance across nanomolar-to-micromolar affinities. Together, IBPC-Kdand the accompanying predictor IDRBindNet provide a quantitative framework for interrogating determinants of IDR recognition and for prioritizing candidate interactions for mechanistic and design studies.