‘All‑to‑all’ world model for biomolecules

Biology runs on molecular fit: proteins, DNA, RNA and small molecules bind like differently shaped keys and locks. ODesign says one artificial intelligence model can design across those parts inside a single system. (arxiv.org) The paper behind the project was posted to arXiv on October 25, 2025, revised on October 28, 2025, and lists 38 authors from Lingang Laboratory, Zhejiang University, Shanghai Jiao Tong University, The Chinese University of Hong Kong, the Massachusetts Institute of Technology and other institutions. (arxiv.org) The team describes ODesign as an “all-atom” model, meaning it represents molecules at the level of individual atoms rather than only amino-acid or nucleotide strings. It says users can pick an epitope, the exact patch on a target molecule they want to hit, and ask the model to generate a matching binder. (arxiv.org) Most generative design tools in this field have focused on one class of molecule at a time, especially proteins. ODesign says it extends that setup to proteins, nucleic acids and small molecules so the same framework can handle designs such as protein-binding RNA or DNA-binding ligands. (odesign1.github.io) The architecture is pitched as a design layer built on the same broad idea that made newer structure predictors useful: one representation for many molecule types. The technical report says AlphaFold3 and RosettaFold-All-Atom showed cross-modality structure prediction was possible, but stopped at predicting existing structures rather than generating new ones for a chosen job. (odesign1.github.io) That distinction matters in drug discovery and synthetic biology, where researchers often want a molecule that does something specific, not just a forecast of how known molecules fold. The ODesign team says earlier protein-centric systems also struggled with the way real cells mix proteins, nucleic acids, metabolites, lipids and metal ions in the same interaction networks. (odesign1.github.io) The authors report gains on protein benchmarks at three levels — whole entities, sequence tokens and atoms — and say the model showed better controllability than narrower baselines. The public project page does not present clinical or wet-lab efficacy claims in the summary, so the headline result is still a modeling and benchmark claim rather than a demonstrated medicine. (arxiv.org) ODesign is not only a paper claim. The team has posted code on GitHub, a technical report, a project page and a web server, and the repository says pre-trained variants are available for different design modes and modalities. (github.com) The immediate test is whether outside groups can turn those benchmark results into molecules that work in the lab. For now, ODesign’s pitch is simple: one model, one interface, many kinds of biomolecular design. (odesign1.github.io)