AlphaFold 3 predicts structures across proteins, DNA, RNA and ligands

Google DeepMind and its commercial spin-off Isomorphic Labs released AlphaFold 3, extending the structure-prediction system beyond proteins to DNA, RNA, ligands and chemical modifications, jointly predicting how such molecules fit together in three dimensions. The developers reported at least a 50% improvement in accuracy on molecular-interaction prediction compared with existing methods, and said the model was the first AI system to surpass physics-based tools on the PoseBusters benchmark for protein-ligand docking.

The accompanying Nature paper broke with the norm set by AlphaFold 2, whose code and weights had been released openly in 2021. AlphaFold 3 was made available only through a free AlphaFold Server web interface, capped at ten prediction jobs per day and restricted to non-commercial use, with no downloadable model. Nature’s editor-in-chief cited “potential implications for biosecurity” among the reasons for waiving the journal’s usual code-sharing requirement; DeepMind’s Pushmeet Kohli separately suggested the restriction protected Isomorphic Labs’ commercial drug-discovery work built on the same model.

The response from structural biologists was sharply critical. An open letter accusing Nature of undermining its own reproducibility standards gathered more than 650 signatures within a week, arguing that a rate-limited server made independent verification of the paper’s claims impractical. DeepMind committed to releasing the underlying code for non-commercial academic use within six months, and did so that November, with model weights available to academics on application — a partial climbdown that left the episode as a marker of how a commercial lab’s release choices about a widely used scientific tool could outlast the science itself as the story.