Accurate structure prediction of biomolecular interactions with AlphaFold 3

What the paper sets out to do

AlphaFold 3 is a structure-prediction model described in a paper published in Nature on 8 May 2024 (volume 630, pages 493–500; DOI 10.1038/s41586-024-07487-w) ^[1]. The paper is about interactions rather than isolated chains: it asks whether a model can predict the structure of a complex formed by a protein and another molecule. Two cases get most of the attention. The first is protein–ligand binding, where the partner is a small molecule sitting in a pocket on the protein. The second is antibody–antigen recognition, where the partners are both proteins and the question is how an antibody's binding site contacts the surface of the molecule it recognises.

Both belong to what is usually called docking. Working out a complex means deciding whether the partners bind at all and, if they do, where each atom ends up relative to the others. The number of possible arrangements is enormous, and the answer depends on how two or more molecules sit together rather than on how a single chain folds, which is what makes the problem harder than predicting one protein on its own.

How the comparison is set up

The paper's evidence is comparative. For protein–ligand interactions the yardstick is docking software: the abstract describes the model's accuracy for these interactions as far greater than that of state-of-the-art docking tools ^[1]. For antibody–antigen prediction the yardstick is AlphaFold-Multimer v.2.3, and the abstract reports substantially higher accuracy than that system ^[1]. Protein–protein interactions are also measured against AlphaFold-Multimer v.2.3 ^[1].

One benchmark is described in detail. The PoseBusters evaluation set used in the paper is made up of 428 protein–ligand structures that were released to the Protein Data Bank (PDB) in 2021 or later ^[1]. Building a test set from recent depositions means the cases are structures that entered the public archive late rather than long-standing entries.

Results

Against classical docking tools such as Vina, the paper reports that the model greatly outperforms them, with Fisher's exact test giving P = 2.27 × 10⁻¹³ ^[1]. Fisher's exact test compares two sets of counts — for example, how many cases each method gets right — and a P value of that size says a gap of the observed size would be very unlikely to appear by chance if the two methods were equally good.

For antibody–antigen prediction the abstract reports substantially higher accuracy than AlphaFold-Multimer v.2.3 ^[1]. For protein–protein interactions, compared with AlphaFold-Multimer v.2.3, a paired Wilcoxon test gave P = 1.8 × 10⁻¹⁸ ^[1]. A paired test of this kind compares the two methods target by target on the same set, so a P value that small indicates that the difference holds across the set rather than being driven by a few cases.

Why it matters

The claims are relative, and the reference points are the tools that have been used for these jobs. A docking program searches for plausible placements and scores them; if a learned model can place ligands and predict antibody–antigen contacts more accurately on standard test sets, the range of questions such predictions can support changes as well. Protein–ligand binding underlies how many small-molecule drugs act, and antibody–antigen contacts are central to how the immune system recognises a target, so better prediction of these complexes is directly relevant to work that depends on knowing their structures.

Authorship and publication record

The paper has 48 authors from Google DeepMind and Isomorphic Labs; two of them also list Stanford University and Princeton University ^[1]. The corresponding authors are Max Jaderberg, Demis Hassabis and John M Jumper ^[1]. It was received on 19 December 2023, accepted on 29 April 2024, and published on 8 May 2024 ^[1].