Google DeepMind and Isomorphic Labs announced AlphaFold 3 on May 8, 2024. The model predicts structures and interactions involving biomolecules, including proteins, DNA, RNA and ligands. The announcement and accompanying Nature paper concern computational predictions, not a completed drug or a guarantee of clinical success. DeepMind announcement
The model extends the scope of prediction beyond isolated protein structures to interactions among different molecule types. The Nature paper describes a simpler pairformer module replacing AlphaFold 2’s Evoformer and a diffusion module predicting atom coordinates. This more precise architecture description distinguishes the paper’s terminology from the announcement’s broader description of an improved Evoformer. Model outputs still need to be interpreted in their research context.
The announcement also describes the AlphaFold Server as a way for scientists to use capabilities for non-commercial research. Access arrangements and usage terms are distinct from the scientific performance claims and should be checked against the current service documentation.
The Nature paper provides the appropriate basis for examining evaluation methods and limitations. A benchmark result applies to a defined test and comparison; it should not be generalized into a promise that every predicted interaction is correct or that drug-discovery experiments are unnecessary. Nature research paper
Archive note: the earlier 2026 launch date and specific accuracy and interaction-count claims were unsupported.