AI for Bio · Model

AF2-Multimer

Multi-chain fine-tune of AF2; the classic protein-complex baseline

Structure & co-foldingBig-tech labUnited KingdomSuperseded by a successor中文

Overview

NameAF2-Multimer · aliases: AlphaFold2-Multimer, AlphaFold-Multimer, AF2-Multimer, AF-Multimer
Full nameAlphaFold2-Multimer
Ambiguous names (not auto-linked)AlphaFold2, AlphaFold
FamilyNot a family entry. Single-model record; the multimer v3 weights of v2.3 are a milestone, not a family member
OrganizationGoogle DeepMind
Organization typeBig-tech lab
CountryUnited Kingdom
Model classCo-folding (protein complexes only)
DomainStructure & co-folding (also: Proteins)
First release · paper revision
Basis: bioRxiv preprint 'Protein complex prediction with AlphaFold-Multimer' v1 source
Other milestones
  • First public evidence: · bioRxiv preprint v1, earliest provable public evidence source
  • Preprint v1: · bioRxiv v1; no journal version of the same paper found source
  • Weights release: · AlphaFold v2.3.0 shipped the multimer v3 weights source
Latest update · new weights
Basis: AlphaFold v2.3.0 released the multimer v3 weights source
The most recent sourced event, not a claim that nothing changed since.
StatusSuperseded by a successor · Complex prediction superseded by AlphaFold 3; kept as a baseline
OpennessCode: Open · Apache-2.0
Weights: Open · CC BY 4.0 (per site record; official source governs)
Code and parameters ship with the AlphaFold repository; licence terms per the official page, text not re-verified this round
Benchmarks reported1 counted (board 1)
Wet-lab validationNone recorded
Unverified fields† benchmarks_reported[0], training_data.a_public, open_status.weights_license, open_status.code_license
As of2026-11-16

What it does

Fine-tunes AF2 on multi-chain data to output protein-complex structures with an interface confidence (ipTM). Proteins only, no ligands or nucleic acids, which is what AlphaFold 3 later superseded; large-scale sampling methods such as MassiveFold use it as their engine.

Benchmarks reported

Every row is one pairing from the survey's relation file; a score belongs to the benchmark page's ranking, where it is shown with metric, eval set, source type and date. 1 of 1 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
22 CAPRIOfficial leaderboard or resultsCASP15-CAPRI official assessment, assessor-run yardstick baselineBoardThe official round assessment lists it as the assessors' yardstick baseline, counted as an official entry; no per-target numbers. †partly verified · evidence

Other evaluations (outside the 44)

  • Seventeen-heterodimer no-template benchmark (Public benchmark not among the 44) Self-reported: Interface accuracy on seventeen heterodimers in a no-template setting; not among the site's 44 benchmarks (developer-reported)
  • Large recent-PDB complex set (Own task set) Self-reported: Developer-built set of 4,446 recent complexes with non-redundant low-template interfaces, compared against AF2 run with a linker (self-reported)

Developers report: interface-prediction success is substantially higher than AlphaFold run with a linker, on both heteromeric and homomeric interfaces, including a no-template setting (self-reported).

Independent evidence: It appears mainly as a third-party baseline: the CASP15-CAPRI assessors used it as their yardstick and the NeuralPLexer3 paper re-ran it as a standard comparison.

Training data

Group in Figure 7: Public data only.

ClassValueDetail
a. Public databasesYesPDB multimers, UniRef and similar public databases; described on the site as general knowledge, not checked item by item source
b. Private or licensed dataNoThe preprint lists public databases only and discloses no licensed data source
c. Own wet-lab data used for trainingNo
d. Wet lab used only for validationNo

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: Google DeepMind is Alphabet's AI lab in London; DeepMind was founded in 2010, acquired by Google in 2014 and merged with Google Brain into Google DeepMind in 2023-04.
  • Funding: Funded from Alphabet's internal budget, no separate rounds
  • Products and partnerships: Successor to AF2 and predecessor of AlphaFold 3 and IsoDDE; MassiveFold runs large-scale sampling on the AF2 line.
  • Related models: AlphaFold 3, MassiveFold, IsoDDE

Reading notes

  1. Its value is as a baseline: a new model's gain over it on complex interfaces is the minimum bar for progress.
  2. Comparison figures in third-party papers are those authors' own re-runs, not developer self-reports, and should not be pooled.
  3. It has no ligand or nucleic-acid capability, so it cannot anchor small-molecule or RNA comparisons.
  4. After 2022-12 only fixes and cleanup commits followed; none counts as a new model generation.

Caveats

  1. It leaves the zero-pairing list this round as an official CAPRI assessment baseline, though that pairing is only medium confidence.
  2. Figures on its own task set and on off-site benchmarks are developer-reported, with no independent replication found.
  3. Training data is listed as general knowledge; per-database sizes are not verified.
  4. Licence terms (code Apache-2.0, parameters CC BY 4.0) follow the older record; the official page governs.

Sources

  1. AlphaFold-Multimer bioRxiv preprint (preprint) ·
  2. AlphaFold repository releases (repository) ·
  3. CASP15-CAPRI official assessment paper (third-party paper)
  4. NeuralPLexer3 paper (third-party comparison) (third-party paper)

Corrections

2 corrections
FieldWasNowBasisFound
benchmarks_reportedZero pairings: no site benchmark pairingCounted as an official CAPRI assessment baseline; leaves the zero-pairing list this roundsource2026-11-16
first_release2021-10 (month only)2021-10-04 (bioRxiv v1)source2026-11-16

Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-16.