Intrinsically disordered proteins and regions (IDRs) lack stable 3D structures, posing challenges for interaction prediction. We present SpatPPI, a geometric deep learning model tailored for IDPPI prediction. SpatPPI leverages structural cues from folded domains to guide the dynamic adjustment of IDRs via geometric modeling, adaptive conformation refinement, and a two-stage decoding mechanism. It captures spatial variability without requiring supervised input and achieves state-of-the-art performance on benchmark datasets. Molecular dynamics simulations further validate its high adaptability to conformational changes in IDRs and strong capacity to generate distinct and structure-aware embeddings. A freely accessible server is available at http://liulab.top/SpatPPI/server .

Conformational dynamics; Geometric deep learning; Intrinsically disordered proteins; Protein–protein interaction; Residue interaction characteristics