AlphaFold 3

What it is

AlphaFold 3 is a structure-prediction model from the Google DeepMind AlphaFold team and Isomorphic Labs, announced on 8 May 2024.^[1] Its subject matter is wider than that of a protein-folding system: it predicts the structure and interactions of proteins, DNA, RNA, ligands (small molecules) and chemical modifications.^[1] The widening matters because much of what a molecule does in a cell depends on what it binds to. A protein that carries a drug, an enzyme reading a stretch of DNA, an RNA folded around a small molecule — each is a complex of several molecule types, and describing such a complex is the problem the model is built for.

How it works

The core of AlphaFold 3 uses an improved version of the Evoformer module that underpinned AlphaFold 2.^[1] The model also uses a Pairformer module.^[2] To turn its internal representation into actual coordinates, it assembles its predictions with a diffusion network, which Google compared to the networks used in AI image generators.^[1][2] The comparison is loose but captures the procedure: the diffusion network starts from a cloud of atoms and converges over many steps on the final structure, instead of placing every atom in a single pass.^[1]

Results

Google reported at least a 50% improvement over existing methods for interactions of proteins with other molecule types, and doubled accuracy for some interaction categories.^[1] On the PoseBusters benchmark, Google reported that AlphaFold 3 was 50% more accurate than the best traditional methods that do not use structural input.^[1] The company also said AlphaFold 3 was the first AI system to surpass physics-based tools for biomolecular structure prediction.^[1] These are the developers' own reports; benchmark results depend on the test set and on what counts as the best existing method, so the figures are best read as evidence of a large step rather than as a permanent ranking.

Access and uses

At launch, scientists could use the model through AlphaFold Server, a free platform for non-commercial research.^[1] Source code and weights were made available for non-commercial scientific use on request in November 2024, and publicly in February 2025, still under the non-commercial restriction.^[2] Isomorphic Labs applies AlphaFold 3 to drug design, both for internal projects and with pharmaceutical partners.^[1] A free server and open weights for research, with applied work routed through a company, is the arrangement under which the model reached its users.

Why it matters

Knowing the shape of a biomolecular complex is usually a prerequisite for reasoning about how it works: which surface a drug might bind to, which parts of two chains touch, why a change in sequence breaks a function. Experimental structure determination is slow and does not succeed for every complex, so a computational method that covers several molecule types at once changes which questions can be asked cheaply. AlphaFold 3's contribution lies in the breadth of molecule types it handles and in the accuracy reported for protein interactions with other molecules, rather than in posing a new kind of question.

Where it sits

Google and Isomorphic Labs published the announcement on 8 May 2024 alongside a Nature paper introducing AlphaFold 3.^[1] The model therefore arrived in two registers at once: a research result described in the scientific literature, and a service that working scientists could use the same day. Its place in the line of AlphaFold systems is defined by that combination — an extension from proteins to complexes of several molecule types, delivered with a benchmark claim and a public access route.