Intrinsically disordered proteins (IDPs) play central roles in signalling, regulation, and disease, yet they remain a severe challenge for current artificial intelligence models of protein structure. Most methods were trained on ordered, single structure data and optimize objectives that implicitly favour compact, well-folded conformations. IDPs, by contrast, populate broad conformational ensembles and display heterogeneous, condition-dependent structures. We argue that IDPs provide a natural benchmark and development ground for the next generation of AI systems, including future artificial general intelligence (AGI) in biology. We summarize the basic biophysics of IDPs and survey existing AI approaches to protein structure and ensemble prediction. Because α-synuclein misfolding and aggregation are central to Parkinson's disease, we focus on α-synuclein, an IDP, as our primary system. We evaluate four ensemble generation pipelines, AlphaFlow, AlphaFlow-MD, AFflecto, and Ensemblify, using contact-probability maps, radii of gyration, and secondary-structure statistics. The analysis reveals systematic tendencies of structure prediction-inspired models to over stabilize compact, helical states, while ensemble aware approaches recover expanded, coil-rich conformations characteristic of disorder. We discuss how such case studies can be turned into quantitative benchmarks for AI models, and outline design principles for future systems that reason natively in terms of ensembles, dynamics, and experimental observables. Finally, we highlight open questions and opportunities at the interface of IDP biology, molecular simulation, and intelligent modelling.
Coskuner‐Weber et al. (2026) studied this question.