AlphaFold 2
Structure prediction · DeepMind · 2021

AlphaFold 2

Predicts the structure of a protein from its sequence, an alignment of its homologues and optional templates. The Evoformer exchanges information between the alignment and residue pairs; a geometric structure module turns those representations into atoms.

In plain language

You give it one protein sequence. It finds thousands of related sequences from other species, notices which positions change together across them, and turns that into a three-dimensional model of the protein.

One way to picture it

Imagine reading the same sentence translated into a thousand languages. Two words that always change together are probably linked in meaning. Two residues that always mutate together are probably touching in space.

Commonly misread as

A high confidence score is not proof that the structure is right, and a low one often means the region has no fixed shape rather than that the model failed.

How it is put together

Select a component to read it
Prepared evidence
Initial MSA and pair state
Outer product mean
Single and pair representations

Try it

Fold a sequence without installing anything
# Open the notebook, paste a sequence, run every cell.
# https://colab.research.google.com/github/sokrypton/ColabFold/blob/main/AlphaFold2.ipynb

# Locally, the same idea with the command line client:
pip install "colabfold[alphafold]"
colabfold_batch my_sequence.fasta out_dir/

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

The CASP14 result in 2020 and the Nature paper in 2021 are the point where computational structure prediction became something experimentalists use by default. The 2024 Nobel Prize in Chemistry cites this work.

02 / What sets it apart

Distinctions

  • The pair representation carries triangle constraints, so geometry is enforced by the architecture and not only learned.
  • Invariant point attention lets the decoder move residues in three dimensions without breaking equivariance.
  • The alignment search, not the network, is often the slowest part of a prediction.
03 / Where it stops

Limits

It predicts one structure per input. Alternative conformations, the effect of a point mutation and the behaviour of disordered regions are outside what a single confident prediction can express.

Weights and code

Checked against the registry, not from memory
RepositorySizeLicenceNote
google-deepmind/alphafoldOpen weights via the reference implementationApache-2.0 code, CC BY 4.0 parameters—
aqlaboratory/openfoldTrainable reimplementationApache-2.0—

Sources

Each number above comes from one of these
Jumper et al. 2021, Nature. Highly accurate protein structure prediction with AlphaFoldPeer reviewedAlphaFold reference implementationCode

Same task, other answers

Structure prediction
Structure prediction2024

AlphaFold 3

One model for the whole complex.

Predicts the joint structure of complexes containing proteins, nucleic acids, small molecules, ions and modified residues.

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