On May 8, 2024, the journal Nature published AlphaFold 3 [1]. It is the latest protein structure model from Google DeepMind and Isomorphic Labs. Nature’s editorial summed up the change. Unlike AlphaFold 2, the new model predicts how proteins interact with other kinds of molecule, including DNA and RNA. In the paper’s words, it predicts the joint structure of complexes of proteins, nucleic acids, small molecules, ions and modified residues.
I think AlphaFold 3 is a remarkable achievement in computer-based biology, but it is easy to read more into it than the paper claims. A predicted structure shows where atoms are likely to be. It does not show that a drug will work in a cell, an animal or a patient, and the paper does not claim that it does.
Why does a protein’s shape matter?
Proteins are the working parts of living things. Their shape is a big clue to what they do. The 2021 AlphaFold 2 paper opened with that point. Understanding a protein’s structure “can facilitate a mechanistic understanding of their function,” the authors wrote [2]. In plain words, knowing the shape helps explain how a protein does its job.
The problem has always been the cost of finding a structure. At the time, scientists had solved around 100,000 unique protein structures by experiment. That is a small fraction of the billions of known protein sequences. A single structure could take months to years of work.
AlphaFold 2 changed that. Its authors described it as the first computer method that could regularly predict protein structures accurately down to the level of single atoms. It could do this even when no similar structure was known.
AlphaFold 3 goes further, because much of what happens in living things depends on molecules fitting together. The authors write that accurate models of complexes are “critical to our understanding of cellular functions and for the rational design of therapeutics” [1]. In other words, these models help us understand cells and plan new medicines.
How well does AlphaFold 3 really work?
The paper makes big claims [1]. For drug-like molecules bound to proteins, the model is far more accurate than the best docking tools. Docking tools are programs that try to fit a small molecule into a protein. For complexes of proteins with DNA or RNA, it beats specialized programs by a wide margin. It also predicts the shapes of antibodies better than the team’s earlier model.
The test with drug-like molecules got the most attention. The team used a benchmark called PoseBusters. It holds 428 protein-ligand structures added to the Protein Data Bank in 2021 or later. Here, a ligand is a small molecule that binds to a protein. The Protein Data Bank is a public archive of solved molecular structures.
The team set a clear rule for counting a prediction as correct. First, they lined up the binding pocket of the predicted structure with the real one. Then they compared the predicted ligand’s atoms with the atoms in the real, measured pose. The typical gap between the predicted and real atoms had to be under 2 angstroms. An angstrom is a tiny unit of length, a fraction of a billionth of a meter.
AlphaFold 3 greatly beat classical docking tools such as Vina. It did this without being given any information from the real structures. The docking tools, in contrast, used information from the solved test structure.
I think that is a strong result. But it is important to see what the test measured. Every score compares a prediction with a structure that someone had already solved by experiment. The team also trained the model on structures from the Protein Data Bank that were found by experiment. The benchmark checks whether the shape is right. It does not ask whether the molecule actually does anything.
What weak spots do the authors admit?
To their credit, the authors give a whole section to limitations. They group them under “stereochemistry, hallucinations, dynamics and accuracy for certain targets” [1].
Stereochemistry is about handedness. Many molecules come in left-handed and right-handed forms, like a pair of gloves. In biology, the difference can matter. The model does not always get this right. When the team ranked predictions, they gave a penalty for wrong handedness. Even so, they saw wrong handedness 4.4% of the time on the benchmark. The model also sometimes places atoms so close together that they overlap.
The second problem is hallucination, which means the model makes up structure that is not really there. AlphaFold 3 is a generative model, a type of AI that produces new outputs from patterns it learned in training. So it “may invent plausible-looking structure even in unstructured regions”. These are parts of a protein that are naturally floppy. The model usually marks such regions with very low confidence. But these regions do not always look loose and ribbon-like, which is how a floppy region usually appears.
Imagine a student who downloads a predicted structure and sees a neat, tightly folded loop. She builds a whole project around a drug binding there. Nobody checked the confidence colors. The loop was floppy all along, and the tidy shape was something the model made up.
The third limitation is the one I think matters most. AlphaFold 3 predicts still structures, not how molecules move in a liquid. The authors give an example with a family of enzymes called E3 ubiquitin ligases. These enzymes take an open shape when nothing is bound to them. They have been seen only in a closed shape when ligands are bound. Yet AlphaFold 3 predicts only the closed shape in both cases. The authors also write that “there are still many targets for which accurate modelling can be challenging”.
Why is a predicted shape only a hypothesis?
Suppose a prediction shows a molecule fitting into a pocket on a protein. It still has to bind in a real test tube. Then it has to bind inside a living cell crowded with other molecules. It has to change what the protein does. It has to reach the right tissue without harming anything else. Finally, it has to make patients better off.
AlphaFold 3 helps only with the first step, and only as a prediction. Everything after that needs real experiments. A good predicted structure can help scientists choose which experiments to run. It cannot replace those experiments.
I have worked on both sides of this question. My doctoral training was in lab-based imaging research, and today I build AI models. So I respect what a model can do, and also what only a measurement can tell you.
The authors seem to agree. They expect structure prediction to keep improving partly because of experiments [1]. Methods such as cryo-electron microscopy will supply new training data. In this way, predictions and experiments help each other improve.
Why did the missing code matter?
In May 2024, the paper also started a second debate, about its code. The paper’s statement on code made two points [1]. AlphaFold 3 would be offered as a web server for non-commercial use, with limits on which ligands and chemical changes were allowed. It also said that “code is not provided”.
Nature’s May 22 editorial noted a contrast. AlphaFold 2’s full code had been open to all researchers. AlphaFold 3 came with pseudocode instead, which is a detailed description of how the model works. The server was free for non-commercial use. But it limited uses such as drug development and capped the number of predictions per day.
According to the same editorial, an open letter to the journal argued that the missing code made the work harder to check and repeat. DeepMind’s Pushmeet Kohli then posted that the team was working to release the model for academic use within six months. The journal said it had not made the decision lightly. It said it wanted its view to start a conversation and not to be the final word. In November 2024, an addendum to the paper announced that the team had released the code for running the model.
Checking a claim is how science earns trust. A tool that others cannot fully inspect is harder to check.
In my view, AlphaFold 3 deserves the excitement. But a predicted structure is a well-informed guess about where atoms sit. The real questions about binding, function and benefit to patients still have to be answered at the lab bench and in the clinic.
So the next time you see a headline about AI and a new drug, look for one detail. Was anything actually tested in a lab, a cell or a person? If not, you are reading about a promising idea that still needs testing. That is still good news. But the work is at an earlier stage than the headline suggests.
References
[1] J. Abramson, J. Adler, J. Dunger, R. Evans, T. Green, A. Pritzel, et al., “Accurate structure prediction of biomolecular interactions with AlphaFold 3,” Nature, vol. 630, no. 8016, pp. 493-500, Jun. 2024, doi: 10.1038/s41586-024-07487-w.
[2] J. Jumper, R. Evans, A. Pritzel, T. Green, M. Figurnov, O. Ronneberger, et al., “Highly accurate protein structure prediction with AlphaFold,” Nature, vol. 596, no. 7873, pp. 583-589, Aug. 2021, doi: 10.1038/s41586-021-03819-2.
Related Articles
