AI-IDP generates experimentally validated conformational ensembles of intrinsically disordered proteins directly from sequence by integrating deep-learning fragment prediction with flexible assembly, revealing that transient α-helices and polyproline-II segments are evolutionarily tuned structural motifs whose local sequence-encoded preferences are sufficient to determine global chain dimensions and whose prevalence patterns across proteomes reflect adaptive functional specialization from prokaryotes to higher eukaryotes.

Intrinsically disordered proteins (IDPs) comprise nearly one-third of the human proteome and play key roles in regulation, signaling, and disease, yet their dynamic nature has resisted accurate structural prediction. Here we introduce AI-IDP, a deep-learning framework that transforms sequence information into experiment-consistent conformational ensembles of disordered proteins. By combining deep-learning-driven fragment prediction with flexible physical assembly, AI-IDP reproduces experimental observables across local, medium-range, and global scales, including transient secondary structure, mutation sensitivity, and overall chain dimensions. Applied to more than 3,000 disordered regions across human and non-human proteomes, AI-IDP reveals that transient α-helices and polyproline-II conformations are pervasive and evolutionarily tuned features of disorder. By uncovering how sequence encodes conformational heterogeneity, AI-IDP provides a practical framework for understanding the structural and functional logic of disordered proteomes and enables rationally targeting the dynamic states that underlie health and disease.