Why do interactions between molecules matter?
In our cells, a protein rarely works alone. It binds to DNA, RNA, other proteins or small molecules. A medicine, for example, often works by fitting into a precise pocket of a protein.
Knowing the 3D shape of these complexes helps us understand a disease and design a drug. But measuring them in the lab is slow and costly. AlphaFold 2, honored by the 2024 Nobel Prize in Chemistry, could estimate the shape of a protein. The next step was doing the same for groups of interacting molecules.
What did the Google DeepMind team change?
Researchers at Google DeepMind and Isomorphic Labs deeply reworked the model. AlphaFold 3 uses a technique called diffusion: it starts from a blurry cloud of points and sharpens it step by step into a precise structure, a bit like a photo coming into focus.
A single model now handles proteins, nucleic acids (DNA and RNA), small molecules, ions and modified proteins. To check it was not simply recalling structures it had seen, the team tested it on structures released after its training cut-off date.
Better than specialist tools
To see how a small molecule binds to a protein, the team used the PoseBusters test, made up of 428 structures. AlphaFold 3 clearly beat classic docking tools such as Vina, even though, unlike those tools, it was not given the protein's shape as a starting point.
It was also more accurate than specialist tools for complexes of proteins with DNA or RNA, and much better than the previous version for antibodies binding their targets. The study shows, for example, a giant complex of 4,665 building blocks (amino acids): the spike protein of a human coronavirus bound to antibodies.
Key figures
- 428protein and small molecule structures in the PoseBusters test, where AlphaFold 3 beats docking tools
- 4,665building blocks (amino acids) in a giant complex: a coronavirus spike protein bound to antibodies
- 1,000model runs used to rank antibody and antigen interactions
From the lab to your care, with Mima
AlphaFold 3 is a lab tool: it does not change your treatment today. In time, it may help design new medicines, but each one will need clinical trials before it reaches patients.
If you wonder what research could mean for you, Mima helps you note a question for your doctor, for example: "Is there a clinical trial that could be right for me?" And she keeps track of the answers.
What this could change for patients, in time
For patients, AlphaFold 3 changes nothing directly today: it is a research tool. But it may speed up the work of labs looking for new medicines, by helping them picture how a molecule will bind to its target before making it.
It may also help explain how an antibody recognizes a virus, or how a protein acts on DNA. This knowledge lays the groundwork for future treatments, but the road to patients is long and goes through clinical trials.
What are the limits of AlphaFold 3?
The authors describe several limits themselves. The model does not always respect chirality, meaning a molecule's handedness, like a left and a right hand. It can also hallucinate, inventing a plausible-looking structure in regions that have none.
It gives a frozen picture, while molecules move all the time inside the cell. And for antibodies, the model has to be run a very large number of times (up to 1,000 runs in the study) to get the best results.
Frequently asked questions
What is the difference between AlphaFold 2 and AlphaFold 3?
AlphaFold 2 mainly estimated the shape of a single protein. AlphaFold 3 models complexes mixing proteins, DNA, RNA, small molecules and ions, with a new diffusion-based design.
Is AlphaFold 3 used to make medicines?
It can help researchers understand how a molecule binds to a protein, one step in drug design. It does not create a treatment on its own.
Is AlphaFold 3 always right?
No. The authors report errors in handedness, structures that are sometimes invented, and frozen shapes even though molecules move.
Who developed AlphaFold 3?
Researchers at Google DeepMind and Isomorphic Labs, in a study published in Nature in 2024.
The original study
Abramson J, Adler J, Dunger J, et al. “Accurate structure prediction of biomolecular interactions with AlphaFold 3” Nature, May 8, 2024. DOI : 10.1038/s41586-024-07487-w
Read the full study in NatureThis article summarizes a scientific study and is not medical advice.