MinedDeep

Protein Complex Prediction with AlphaFold-Multimer

2 min read

Protein Complex Prediction with AlphaFold-Multimer

Type: paper Slug: protein-complex-prediction-with-alphafold-multimer—hassabis Sources: protein-complex-prediction-with-alphafold-multimer—hassabis Last updated: 2026-05-13

Summary

Evans, O’Neill, Pritzel, Antropova, Senior, Green, Žídek, Bates, Blackwell, Yim, Ronneberger, Bodenstein, Zielinski, Bridgland, Potapenko, Cowie, Tunyasuvunakool, Jain, Clancy, Kohli, Jumper, and Hassabis (2021) introduced AlphaFold-Multimer, an extension of AlphaFold2 that predicts the structures of protein complexes (multi-chain assemblies). The system modifies the AlphaFold2 architecture to handle multiple protein chains, predicting both intra-chain and inter-chain contacts to produce full complex structures. This addressed a major limitation of the original AlphaFold2.

Core content

Problem: AlphaFold2 predicts single-chain protein structures with high accuracy but cannot natively predict how multiple proteins assemble into complexes. Most biological functions involve protein complexes, not isolated chains (paper—protein-complex-prediction-with-alphafold-multimer §Introduction).

Key modifications to AlphaFold2:

  • Input representation: Extended to handle multiple sequences (one per chain) with explicit chain identifiers in the MSA and pair representations (paper—protein-complex-prediction-with-alphafold-multimer §Methods).
  • Pair representation: Includes inter-chain residue pairs (not just intra-chain), allowing the model to predict chain-chain interactions (paper—protein-complex-prediction-with-alphafold-multimer §Methods).
  • Training data: Curated set of experimentally determined protein complex structures from the PDB (paper—protein-complex-prediction-with-alphafold-multimer §Methods).
  • Assembly module: Predicts stoichiometry and chain ordering for heteromeric complexes (paper—protein-complex-prediction-with-alphafold-multimer §Methods).

Results:

  • Achieved higher accuracy than dedicated docking methods on heterodimer benchmarks (paper—protein-complex-prediction-with-alphafold-multimer §Results).
  • Successfully predicted structures of complexes with up to ~10 chains (paper—protein-complex-prediction-with-alphafold-multimer §Results).
  • Performance was strongest for complexes where good MSAs were available; degraded for complexes with few homologs (paper—protein-complex-prediction-with-alphafold-multimer §Results).

Connections- Theme: theme—protein-folding, structural-biology

  • Project: AlphaFold-Multimer
  • Collaborators: Richard Evans (co-first), Michael O’Neill (co-first), Alexander Pritzel (co-first), Natasha Antropova (co-first), John Jumper (co-first), Andrew Senior, Tim Green, Pushmeet Kohli
  • Era: alphafold-era
  • Venue: venue—bioRxiv (preprint, not peer-reviewed)
  • Extends: paper—highly-accurate-protein-structure-prediction-with---hassabis — AlphaFold-Multimer builds on AlphaFold2

Honest Gaps

  • Metadata lists only 4 co-authors; the actual paper has ~22 authors.
  • This is a bioRxiv preprint — it has not undergone formal peer review.
  • Performance on large complexes (>5 chains) and transient interactions is limited.
  • The system cannot predict complexes involving nucleic acids, lipids, or small molecules.
  • Assembly prediction (which chains form a complex and in what stoichiometry) remains a challenge — the system requires the input chains to be specified.
  • Like AlphaFold2, performance depends heavily on MSA depth — complexes with few evolutionary homologs are poorly predicted.

Graph View

Backlinks