Structure prediction · DeepMind · 2024

AlphaFold 3

Predicts the joint structure of complexes containing proteins, nucleic acids, small molecules, ions and modified residues. A Pairformer reasons over tokens and pairs, and a diffusion module generates atom coordinates.

In plain language

The same job as AlphaFold 2, extended to everything that sits next to a protein: DNA, RNA, a drug molecule, a metal ion. It writes the whole assembly at once instead of predicting the protein and docking the rest afterwards.

One way to picture it

AlphaFold 2 drew the building. AlphaFold 3 draws the building with the furniture already in it, and it sketches the same room several times so you can see which arrangement it keeps returning to.

Commonly misread as

Because the decoder generates rather than measures, it can draw a tidy pose for a ligand that does not really bind. The confidence ranking exists for that reason and should be read, not skipped.

How it is put together

Select a component to read it

Token and pair state

Pair representation

Conditioning

Try it

Run an open reimplementation

git clone https://github.com/bytedance/Protenix
cd Protenix && pip install -e .

# One YAML per job: chains, ligands, and what to sample.
protenix predict --input examples/example.json --out_dir ./output --seeds 101

These snippets have not been executed here. Versions move; check the model card before trusting a line of it.

01 / Why it is here

Standing

Reported accuracy above specialised docking tools for protein and ligand interactions, above nucleic-acid-specific predictors, and above AlphaFold-Multimer 2.3 for antibody and antigen pairs, within one framework.

02 / What sets it apart

Distinctions

The MSA representation leaves the deep trunk. The Pairformer keeps the triangle updates and drops the alignment stream.
The deterministic structure module is replaced by diffusion over atoms, so sampling returns a set of structures.
Ligands and modified residues are tokenised per atom instead of being absent from the representation.

03 / Where it stops

Limits

A generative decoder can produce plausible geometry for regions that have none, which is why confidence ranking is part of the method. Open reimplementations exist because the original weights are not freely redistributable.

Weights and code

Checked against the registry, not from memory

Repository	Size	Licence	Note
bytedance/Protenix	368M in Protenix-v1	Open source	Independent reimplementation of the AlphaFold3 approach
boltzgen/boltzgen-1	Open weights	MIT	Related open all-atom generative model

Sources

Each number above comes from one of these

Abramson et al. 2024, Nature. Accurate structure prediction of biomolecular interactions with AlphaFold 3Peer reviewed Protenix-v1 preprint, 2026Preprint

Same task, other answers

Structure prediction

Structure prediction2021

AlphaFold 2

Evolution in, coordinates out.

Predicts the structure of a protein from its sequence, an alignment of its homologues and optional templates.

Explore the architecture Structure prediction2023

ESMFold

Fold without searching for relatives.

Predicts a structure from a single sequence with no alignment search at query time.

Explore the architecture Structure prediction2026

ESMFold2

A language model encoder with an all-atom diffusion decoder.

Predicts all-atom structures of proteins and their complexes from ESM C representations, with an optional alignment for difficult targets.

Read the entry

Back to the catalogueCompare it with its neighbours