HyperIDR: a multi-scale semantic hypernetwork for identification of intrinsically disordered regions
Abstract
Background Intrinsically disordered regions (IDRs) are widespread in proteins and play essential roles in regulation, signaling, and molecular recognition. IDRs of different lengths often exhibit distinct sequence and functional properties, and proteins containing intrinsic disorder may harbor diverse combinations of short and long IDRs. Although several length-dependent predictors have shown promising results, they generally rely on explicit sequence partitioning or separate models for different disorder lengths, limiting their ability to capture coexisting disordered regions of different lengths within individual proteins. Results In this study, we propose a multi-scale semantic hypernetwork, named HyperIDR, for IDR prediction. HyperIDR integrates protein language model embeddings with a semantic prompt routing module to adaptively select scale-relevant sequence cues. These cues guide a lightweight hypernetwork to generate short-, general-, and long-scale representations, which are further integrated through residue-level gating without explicit sequence partitioning. This design enables HyperIDR to model coexisting disorder regions of different lengths in a unified framework. Experimental results on five independent test datasets show that HyperIDR achieves robust predictive performance and maintains stable accuracy across proteins with diverse short- and long-IDR compositions. Mechanistic analyses further demonstrate that the three scale-specific pathways capture complementary information, including local disorder-related patterns, extended disorder-related regions, and sequence representations associated with ordered structural contexts. Conclusions We developed HyperIDR as a unified multi-scale framework for modeling heterogeneous IDR compositions within proteins. The experimental results suggest that adaptive integration of scale-specific representations improves the robustness and biological interpretability of IDR prediction. Finally, a web server is accessible at http://bliulab.net/HyperIDR/ .