IDPForge models ensembles of flexible proteins and preserves folded domains
Published in Artificial Intelligence, Bioinformatics, Journals, Tools.
Published November 1, 2026.
A Nature Communications paper published on October 30 presents IDPForge, a machine-learning tool for modeling proteins whose flexible regions adopt many configurations rather than one stable structure.
De Castro and colleagues use a transformer-based diffusion model to generate ensembles of structures with individual atoms represented. The method handles fully disordered proteins and proteins combining flexible regions with folded domains, preserving those folded components.
The generated ensembles showed agreement with experimental nuclear magnetic resonance and small-angle X-ray scattering data. Users can optionally guide sampling with experimental constraints. The method does not require training tailored to each individual sequence.
The research includes collaborators at SickKids and the University of Toronto. Its open-source implementation could support structural biology, but agreement with averaged experimental measurements does not prove that every generated conformation occurs in nature or establish drug-discovery success.
This is the peer-reviewed publication of work previously circulated as a March 2026 preprint. The accepted article remains subject to editorial changes. The authors report no competing interests.
Source: De Castro et al., Nature Communications, October 30, 2026.