AlphaFold 3 Paper and Access Record

AlphaFold 3 Paper and Access Record

The May 2024 Nature paper describes a model for predicting joint structures of complexes that can include proteins, nucleic acids, small molecules, ions, and modified residues. The system uses a generative diffusion process and reports results across interface-specific benchmark categories.

The protein-ligand result frequently summarized as outperforming physics-based tools is tied to the PoseBusters comparison and its stated information conditions. It is not evidence that AlphaFold 3 replaces every physics-based method, predicts molecular dynamics, validates a drug, or produces a clinical outcome.

The paper reports confidence measures, training cutoffs, test-set construction, baseline conditions, sample generation, and selection rules. Those details belong beside the benchmark result because they determine what was measured.

At launch, AlphaFold Server offered non-commercial research access. The launch did not make the entire model open source. Google DeepMind and Isomorphic Labs later released model code and weights for academic use in November 2024. Current DeepMind materials route readers to the server, database, and academic download.

Google DeepMind and Isomorphic Labs co-developed the model. Their launch and company descriptions are useful for role and access history, but their impact and drug-discovery claims remain vendor statements unless supported by independent evidence.