What happened?
AlphaFold 3 is an AI model that predicts the 3D shape of groups of molecules joined together, such as a protein holding a strand of DNA or a small drug-like molecule Google blog ↗. It was announced in May 2024 with a peer reviewed paper in Nature Google blog ↗ Nature paper ↗.
AlphaFold 2 predicted the shapes of single proteins. But in cells, proteins work by binding to other things: DNA, RNA (a working copy of genetic information), ions, and small molecules called ligands, which include many medicines Google blog ↗. AlphaFold 3 takes a list of these molecules and predicts how they sit together in 3D Nature paper ↗.
The developers report at least a 50% improvement over existing methods for protein interactions with other molecule types Google blog ↗. On a test set called PoseBusters, which checks how drug-like molecules sit in proteins, they say it was 50% more accurate than the best traditional physics based tools Google blog ↗. At launch the code was not released, which led to public criticism from scientists Undark opinion ↗. Six months later, in November 2024, DeepMind made the code downloadable for non commercial use, with model weights (the trained settings) available on request to academics Nature News ↗.
What are the three pieces?
The model
An updated version of AlphaFold 2's core plus a diffusion network, the same family of method used in AI image generators Google blog ↗.
AlphaFold Server
A free website for non commercial research where scientists can model complexes without coding or big computers Google blog ↗.
Isomorphic Labs
A sister company that uses AlphaFold 3 with its own models on drug design projects and with pharmaceutical partners Google blog ↗.
Drugs work by fitting into proteins, so a single tool that predicts how many kinds of molecules fit together could help scientists choose which ideas to test first.
Leapscope interpretation of the reported result.How did AI help?
The AI model is the breakthrough itself. Researchers at Google DeepMind and Isomorphic Labs designed and trained it, and the Nature paper compares its predictions with structures that scientists had measured in the lab Nature paper ↗. The paper reports it is far more accurate than standard docking tools for proteins with small molecules, and better than earlier tools for protein and DNA or RNA pairs and for antibodies binding their targets Nature paper ↗.
Improvement figure is the developers' claim Google blog ↗; error rate from the Nature paper's limitations section Nature paper ↗.
The authors list clear limits Nature paper ↗. Because it is a generative model, it can invent plausible looking structure in floppy regions of a protein, a problem called hallucination. It sometimes gets the mirror image form of a molecule wrong, in 4.4% of cases on one benchmark. It predicts one still picture, not how molecules move in a living cell, and it can pick the wrong shape when a protein has more than one Nature paper ↗.
Which fields could this affect?
Some uses are available now and others are possible future value; the connections below are our assessment.
Structural biology
Researchers can model proteins with DNA, RNA, ions and some ligands through the free server Google blog ↗. This helps them form ideas before slow and costly lab work.
Explore scienceEarly drug research
Isomorphic Labs is using the model on internal drug design and partner projects Google blog ↗. Results from those projects have not been shown in the sources we reviewed.
Explore healthcareAntibody and vaccine design
The paper reports better antibody and target predictions than the previous AlphaFold version Nature paper ↗. Accuracy improves when many runs are made, which adds computing cost Nature paper ↗.
Explore healthcareApproved medicines
A predicted fit does not show that a drug is safe or works in people. No approved medicine from AlphaFold 3 appears in the sources we reviewed.
What has been checked?
The evidence is a peer reviewed Nature paper with benchmark tests, plus a public server and later code release. Leapscope reviewed these sources; we did not repeat the experiments.
Shown so far
- The Nature paper reports higher accuracy than docking tools and earlier specialised predictors on several benchmark sets Nature paper ↗.
- A free server for non commercial research launched with the model Google blog ↗.
- Inference code was released in November 2024, with weights on request for academic users Nature News ↗.
Still unknown
- How well independent groups reproduce the reported accuracy now that the code is out.
- How often predictions for new drug-like molecules hold up in lab tests.
- Whether its use at Isomorphic Labs leads to medicines that reach patients.
Evidence status: Published research. Stage: Usable. Available through a free server and code for non commercial research.
From prediction to medicine
This is our suggested way to follow it, not a promised timetable.
Can I use it today?
Researchers can use the free AlphaFold Server for non commercial work, and academics can download the code and request the model weights Google blog ↗ Nature News ↗ GitHub ↗. It is a research tool, not a medical product, and its predictions still need lab checks.
A few things you might be wondering
Is AlphaFold 3 open source?
Not fully. The code can be downloaded for non commercial use, but the model weights are only available on request to scientists with an academic affiliation Nature News ↗.
Did the AI design a new drug?
No. It predicts how molecules fit together Google blog ↗. Choosing, making and testing a drug is still done by scientists, and a good fit does not prove a drug is safe or effective.
Why did some scientists criticise the launch?
The Nature paper first appeared without its code, which critics said made results hard to check and reproduce Undark opinion ↗. DeepMind released the code about six months later Nature News ↗.
Go straight to the sources
Checked Oct 8, 2026. The first source is the original announcement or research. Later sources add independent context; background pages do not validate the result on their own.
01The announcement describing the model, accuracy claims, AlphaFold Server and Isomorphic Labs' role.
The peer reviewed paper with benchmark results and a detailed limitations section.
Report on the code release for non commercial use and weights on request for academics.
An outside researcher's argument that releasing the paper without code undermined reproducibility.
The released inference code and the terms for using the model parameters.