Diffusion models take on every molecule of life: AlphaFold 3

Protein folding was solved, more or less, by AlphaFold 2. The harder problem that remained was predicting not just what a protein looks like in isolation, but what it looks like when it binds to the things it actually does: a strand of DNA, a turn of RNA, a small-molecule drug, a metal ion, a post-translationally modified residue. These interactions are where biology happens and where drugs are designed, and no model before 2024 could predict them all in one system. The paper that arrived on 8 May 2024, published in Nature and signed by a large team from Google DeepMind and Isomorphic Labs, is AlphaFold 3, and it tackles the whole molecular zoo at once.

The architectural shift from AlphaFold 2 to AlphaFold 3 is as significant as the leap from the first version to the second. AlphaFold 2 used a structure module built on invariant point attention, an equivariant network that refined a 3D structure from the pair representations produced by the Evoformer trunk. AlphaFold 3 replaces that module with a diffusion model. The trunk still produces pair-level and token representations, but the actual coordinates are generated by a diffusion process that steps from random atomic positions toward a coherent structure, conditioned on the trunk’s embeddings. The same denoiser handles every atom type, whether it belongs to a backbone, a nucleic acid base, a ligand or an ion.

The benchmarks are the story. On protein-ligand interactions, AlphaFold 3 improves by at least 50 percent over the best classical docking tools across a broad benchmark set, and for some categories it doubles the accuracy. On protein-nucleic acid complexes, it outperforms tools built specifically for DNA and RNA binding. On protein-protein complexes, it matches or exceeds the specialized competitors. The model also predicts covalent modifications, glycans and missing atoms, which means the output is closer to the kind of structure a crystallographer would deposit than a stripped-back polypeptide trace.

What makes the diffusion architecture work is that it treats the whole complex as a single generative problem. Earlier systems typically predicted a protein fold and then docked a ligand or a nucleic acid into it as a separate step, which locked in errors from the first stage. AlphaFold 3 refines all atoms jointly, so a ligand’s pose and a protein’s side-chain conformations are optimized against each other in the same process. The team also introduces a memory-efficient attention mechanism that lets the model handle complexes of up to several thousand tokens without the quadratic cost that would otherwise be prohibitive.

The caveats are as important as the gains. AlphaFold 3 predicts a single static structure, not an ensemble, and it does not yet capture conformational flexibility or allosteric transitions. Its confidence estimates are well-calibrated for the categories it was trained on, but novel folds or unusual ligand scaffolds still carry larger errors. The model is also closed-source in its weights at launch, though the code is available for inference and the team has emphasized that the research use case is unencumbered.

A year after publication, the practical effect has been visible in the number of drug discovery teams running AlphaFold 3 poses as a first-pass filter. The model has not replaced crystallography or cryo-electron microscopy, and it has not solved the harder problem of predicting binding affinity. What it has done is collapse the cost of getting a first structural guess for a complex from months to minutes. That is the kind of change that quietly reshapes what experiments get proposed.