AI for Bio · Model

AlphaFold 3

Biomolecular co-folding model that later co-folding models use as the reference baseline

Structure & co-foldingBig-tech labUnited KingdomActiveSite entity page (Chinese site)中文

Overview

NameAlphaFold 3 · aliases: AlphaFold 3, AlphaFold3, AF3
Full nameNo expanded name; official materials use AlphaFold 3
Ambiguous names (not auto-linked)AlphaFold, AlphaFold Server
FamilyNot a family entry. Single-model record; the earlier complex model AF2-Multimer has its own record
OrganizationGoogle DeepMind, Isomorphic Labs
Organization typeBig-tech lab
CountryUnited Kingdom
Model classCo-folding (protein / nucleic acid / small molecule / ion)
DomainStructure & co-folding (also: Small molecules)
First release† partly verified · journal
Basis: Journal paper online with the official server the same day; code and weights followed with v3.0.0 on 2024-11-11
Note: No primary metadata for this date was saved this round, so it is marked unverified source
Other milestones
  • Journal online: † · Journal paper online source
  • Code release: · Code public with GitHub v3.0.0 source
  • Weights release: † · Weights available on request with v3.0.0 source
  • Official announcement: † · Official server launch announcement source
Latest update · code release
Basis: GitHub release v3.0.5
v3.0.5 speeds up search and MSA featurisation and adds a CPU flag and a shard helper; v3.0.3 (2026-06-09) relicensed the code to Apache-2.0 source
The most recent sourced event, not a claim that nothing changed since.
StatusActive · Code releases within 12 months, no new generation
OpennessCode: Partly open · Apache-2.0 from v3.0.3 (2026-06-09)
Weights: On request · on request, non-commercial per earlier note, not re-verified
Licence change: · code release
Basis: v3.0.3 release notes relicensed the code to Apache-2.0 source
Code and weights terms must be stated separately: code is Apache-2.0, weights are on request and their terms are still unconfirmed (unverified)
Benchmarks reported6 counted (board 3, self 1, retrospective 2, cited only 2, unconfirmed 2)
Wet-lab validationNone recorded
Unverified fields† first_release, open_status.weights_license, release_milestones.weights_release, benchmarks_reported[2], benchmarks_reported[4], benchmarks_reported[9]
As of2026-11-16

What it does

A diffusion module generates all-atom coordinates directly, covering protein-ligand, protein-nucleic-acid and antibody-antigen complexes in one model. It is the de facto co-folding standard: open efforts such as Boltz-1/1x, Chai-1, Protenix and OpenFold3 set out to reproduce or approach it.

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. 6 of 10 pairings are counted.

BenchmarkRelationEval set / protocolSource typeNote and evidence
15 CAMEOOfficial leaderboard or resultsoperator-run complex and ligand baseline, 2024-01 to 2025-06BoardCAMEO operators ran AF3 as a baseline (2025 assessment paper); maintainer-run, not a developer submission. †partly verified · evidence
31 FoldBenchOfficial leaderboard or resultsmaintainer leaderboardBoardMaintainer leaderboard (FoldBench repo) · evidence
34 AbBiBenchOfficial leaderboard or resultsregistered interface-score entryBoardBenchmark-registered ipSAE correlation (not structure accuracy), read from paper figure, not rerun †unverified · no source link
05 PoseBustersDeveloper report, official protocolPoseBusters V1 (428), official setSelf-reportedSelf-report on the official V1 set; V2 numbers are third-party recomputations · evidence
01 CASPRetrospective dataset useCASP15 public RNA targetsThird-partyAF3 paper used CASP15 RNA targets post hoc; in CASP16 AF3 is only a component of other groups, no own entry · evidence
33 PoseXRetrospective dataset useself-docking / cross-docking, benchmark authors' runThird-partyPoseX authors ran AF3 under the same protocol (third party) †partly verified · evidence
25 RNA-PuzzlesCited as comparison only (not counted)CASP15 RNA, quoted in another paperThird-partyCASP15 RNA value from its own paper, quoted by NeuralPLexer3; not RNA-Puzzles proper †partly verified · evidence
30 PLINDER / PINDER / Runs N' PosesCited as comparison only (not counted)Runs N' Poses, best@5Third-partyNumbers quoted second-hand from the Pearl paper †partly verified · evidence
12 DUD-E / LIT-PCBAUnconfirmed (not counted)—Unconfirmed sourceOnly a second-hand pose claim in an RSC 2026 paper; no numbers; the benchmark paper does not evaluate AF3 †disputed · no source link
22 CAPRIUnconfirmed (not counted)CASP16 complexesUnconfirmed sourceNo standalone entry appears in the official CAPRI assessment material; only secondary sources call it a component of most pipelines. †disputed · evidence

Other evaluations (outside the 44)

  • Third-party prospective crystal-structure evaluation (third party prospective evaluation) Third-party: A prospective comparison on about 551-557 newly solved complexes: good pose quality, but confidence is not fit for virtual-screening ranking (preprint plus second-hand)
  • Role as a control in 2026 boards (Comparison against baselines only) Third-party: Most new models' papers list it as the main control, while the developers rarely report new evaluations

Developers report: protein-ligand docking is significantly better than classical docking tools, with protein, nucleic acid, small-molecule and ion complexes unified in one model.

Independent evidence: Several third-party papers recomputed it under stricter protocols and a maintainer leaderboard includes it; leakage-controlled splits show accuracy falling as similarity drops, indicating memorisation.

Training data

Group in Figure 7: Public data only.

ClassValueDetail
a. Public databasesYesPublic PDB plus distillation data; the paper gives no per-database sizes source
b. Private or licensed dataNot disclosed
c. Own wet-lab data used for trainingNo
d. Wet lab used only for validationNoThe developers report no wet-lab work

Wet-lab validation

None. The developers report no wet-lab validation.

Background

  • Organization: The model was built by a big-tech AI lab together with its sister company, which was spun out of the same parent in 2021, is headquartered in London and positions itself as an AI drug-design company; it reports over 350 staff across London, Cambridge (Massachusetts) and Lausanne (2026-05 reporting).
  • Funding: 2025-03-31 first external round of $600M led by Thrive Capital with GV and Alphabet; 2026-05-12 a $2.1B Series B again led by Thrive Capital, adding MGX, Temasek, CapitalG and a UK sovereign AI fund; valuation undisclosed.
  • Products and partnerships: 2024-01 deals with two large pharma companies totalling $82.5M upfront and up to about $2.9B in milestones, excluding royalties. The internal engine IsoDDE (disclosed 2026-02-10) extends it; the official blog gives no numbers page.
  • Related models: AF2-Multimer, OpenFold3, Protenix, Boltz-1 / Boltz-1x, Chai-1, IsoDDE

Reading notes

  1. Its docking results differ widely across papers because PoseBusters V1 and V2, sample counts and pose-selection rules differ; the numbers are not comparable across papers.
  2. All co-folding models degrade once results are stratified by similarity to training data, so its high scores should be read per similarity bin.
  3. In third-party prospective evaluations pose quality is good, but confidence is not sufficient for virtual-screening ranking on its own.
  4. The code is Apache-2.0 while the weights remain on request with unconfirmed terms, so commercial usability turns on the weights side.

Caveats

  1. Code and weights terms are separate: the weights terms are unconfirmed and must not be inferred from the code licence (unverified).
  2. The first-release date lacks primary metadata and is marked partially verified.
  3. Two pairings rest only on the site's own pages and must display as unverified.
  4. Challenges whose official assessments contain no entry of its own are not counted as established relations here.

Sources

  1. The model's original journal paper (journal)
  2. Paper full text (public archive) (paper)
  3. Official repository releases (repository)
  4. One of the third-party recomputation papers (third-party paper)
  5. A second third-party recomputation paper (third-party paper)
  6. Baseline paper by the continuous-evaluation operators (third-party paper)
  7. Maintainer leaderboard repository (leaderboard)
  8. Re-run paper by the benchmark's authors (third-party paper)
  9. Official announcement of the first external round (official site)

Corrections

3 corrections
FieldWasNowBasisFound
open_statusCode and weights are both limited to non-commercial academic requestCode is Apache-2.0 from v3.0.3 (2026-06-09); weights are on request with their own, still unconfirmed termssource2026-11-16
first_releaseThe first release was treated as fully verified from the journal dateThe date stands, but with no primary metadata this round it is downgraded to partially verifiedsource2026-11-16
benchmarks_reportedThird-party recomputations and developer self-reports were merged into one cellSplit across the seven relation types: official board, self-report on the official protocol, third-party re-run and unconfirmed each stand aloneno source link2026-11-16

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