Model type
Pairformer plus diffusion model for joint biomolecular structure prediction.
Configuration
AlphaFold3 (MSA)
AlphaFold 3 predicts three-dimensional structures of complexes containing proteins, nucleic acids and other molecular components. It combines a Pairformer representation network with an atomic-coordinate diffusion model. This entry describes the model and local implementation; the hosted AlphaFold Server has a separate profile.
Sources (2)
AlphaFold 3 paper; AlphaFold 3 README · Abstract; Model architecture; README: Installation and UsageHow it worksAlphaFold 3 architecture
Conceptual architecture based on the paper. Exact preprocessing and sampling settings belong to each evaluation.
Sources
AlphaFold 3 paper · Fig. 1d; Fig. 2; Model architectureOverview
Inputs
Protein, DNA and RNA sequences; chemical components; optional MSAs and structural templates.
Sources
AlphaFold 3 input specification · Top-level structure; protein, RNA, DNA and ligand inputsOutputs
Predicted structures in mmCIF plus confidence outputs, including pLDDT, PAE, pTM and ipTM.
Sources
AlphaFold 3 output specification · Output directory structure; confidence outputsAccess
Public inference implementation; weights downloaded directly from Google under separate non-commercial terms.
Sources
AlphaFold 3 README · Obtaining Model Parameters; Installation and UsageSource reviewed · Automated source review, 2026-09-16. All specifications and missing details
Evaluations and results
1 evaluation · 1 metric rows. Different protocols are not a single leaderboard.
Filter evaluations
Applied filters: All linked evaluations
| Tested configuration | Protocol and dataset | Finding | Evidence and details |
|---|---|---|---|
| Configuration: AlphaFold3 (MSA) | Protocol: ESMFold2 Runs N’ Poses reported comparison msa: Runs N’ Poses ligand pass rate (MSA) Dataset subset: Runs N’ Poses complete-case intersection: 2,573 scored ligands (ESMFold2 Runs N’ Poses reported comparison split) | 69% ligand_pass_rate percent · higher Uncertainty: Not reported Coverage: unit: ligands; scored: 2573; eligible: unreported; note: Complete-case intersection; source 2600 systems is not a ligand denominator. | Author-reported evaluation · source checkedMethods, coverage and sourceRuns N’ Poses receptor–ligand co-folding; source benchmark 2,600 systems. Figure 2C reports n=2,573 scored ligands on the intersection where all models produced valid predictions, after excluding undefined SuCOS scores. Multiple ligands in one system are scored independently. Five seeds × five diffusion samples per target; select top candidate by ipTM. Success requires lDDT-PLI >0.8 and BiSyRMSD <2 angstrom. Baselines use 10 recycles and 200 diffusion steps; ESMFold2 uses 10 or 20 loops as labelled and truncated 68-step diffusion. Single-sequence and MSA conditions remain separate. Aggregation: Not reported ESMFold2 primary paper v1, Figure 2C Runs N’ Poses · PDF page 5, Figure 2C, Runs N’ Poses subpanel (right), msa block, bar 8 from left (AlphaFold3 (MSA)), exact printed bar label |
Source checking is not independent reproduction. Release 2026-09-23-2b89723c6dd9.
Use this model
How it works, versions and access
Related profile: AlphaFold 3. This page retains the exact record and its evaluation context.
This configuration
Author-evaluated folding configuration with conditioning and loop count retained from Figure 2C.
- record
- AlphaFold3 (MSA)
- configuration
- Not reported
- entity type
- Configuration
How it works
Joint structure prediction
Sequence, chemical and evolutionary features feed a Pairformer, which builds representations of individual tokens and their relationships. A diffusion module then predicts atomic coordinates. Separate heads estimate confidence. The paper describes 48 Pairformer blocks; the architecture models complexes jointly rather than treating every partner as a separately folded structure.
Sources
AlphaFold 3 paper · Main text: Model architecture; Fig. 1d and Fig. 2Training and evaluation context
The standard model uses a structural training cutoff of 30 September 2021. The dedicated PoseBusters Methods section reports a separate model with a 30 September 2019 cutoff, although other training passages disagree (see limitations). Model seeds, templates, input information and ranking also affect the reported comparison; paper evaluation variants are not automatically identical to current downloadable weights.
Sources (2)
AlphaFold 3 paper; AlphaFold 3 supplementary information · Main paper Methods: Training regime, Inference regime and PoseBusters; supplement Section 5.2Local access
The pinned repository provides inference code and a direct Google-hosted weights download. Its README is more current on access than the server FAQ, which still describes an application form. Code and model parameters have different licences; the server output terms should not be substituted for the local weights terms.
Sources (4)
AlphaFold 3 README; AlphaFold 3 code licence; AlphaFold 3 weights terms; AlphaFold Server FAQ · README: Obtaining Model Parameters and Licences; LICENSE; weights terms: Key things to know; FAQ: model accessTraining data and provenance
Training combines experimental PDB structures with approximately 41 million predicted protein monomers, about 25,000 disorder-focused protein complexes and about 65,000 predicted RNA structures. The supplement also lists transcription-factor examples used during fine-tuning. Its sequence-search resources include UniRef90, UniProt, BFD/Uniclust30, MGnify, Rfam and RNAcentral. These resources have different versions and dates: the structural training cutoff is not a cutoff for every sequence database.
Sources
AlphaFold 3 supplementary information · Sections 2.2 and 2.5; Table 3; Section 2.5.2 distillation datasets (PDF pages 8–9, printed pages 3–4)Strengths, limitations and unresolved questions
Strengths and limitations
Strengths supported by sources
- One architecture handles several molecular component types and their joint structures. This is a capability description, not evidence that every complex will be accurate.
Sources
AlphaFold 3 paper · Abstract; Model architecture - The local implementation exposes input, template and output specifications, allowing an evaluation configuration to be documented.
Sources
AlphaFold 3 input specification · Input format and optional input fields
Limitations and conditions
- Predictions can contain incorrect chirality, atomic clashes or spurious structure in disordered regions. Confidence and structural plausibility need separate inspection.
Sources
AlphaFold 3 paper · Model limitations; Fig. 5 - Sampled structures are not a calibrated solution-state ensemble. Prediction confidence does not establish binding affinity or experimental function.
Sources
AlphaFold 3 paper · Model limitations: dynamics and conformational states; confidence outputs are structure-quality estimates - The public code licence does not remove the separate non-commercial restrictions on weights and outputs.
Sources
AlphaFold 3 weights terms · Key things to know - The source is internally inconsistent about the PoseBusters training cutoff. Its dedicated PoseBusters Methods section and Results specify 2019-09-30, while the general Training regime and supplement Section 5.2 say 2021-09-30. The separate evaluation variant is retained; this profile does not resolve the discrepancy or assign its result to a current downloadable checkpoint.
Sources (2)
AlphaFold 3 paper; AlphaFold 3 supplementary information · Main paper Results and Methods: PoseBusters versus Training regime; supplement Section 5.2, printed page 29
Profile review details
Read primary paper XML, pinned official repository documentation and licences. Reviewed public server FAQ separately. No model run, independent performance replication or human review. Supplementary PDF reviewed, including visual checks of Tables 3 and 6. Conflicting cutoff statements remain explicit. A second automated reviewer checked the AlphaFold source claims and service/model distinction; this is not human review or experimental reproduction.
Stable record: discovery-model-alphafold-3Specifications
Inputs, training, access and other details
Explanatory profile: source reviewed · Automated source review, 2026-09-16. Review applies to the cited claims; unresolved fields are listed below. Numerical results retain their own review status.
| Property | Description and evidence |
|---|---|
| Model type | Pairformer plus diffusion model for joint biomolecular structure prediction.SourcesAlphaFold 3 paper · Model architecture; Fig. 1 |
| Architecture | A 48-block Pairformer builds token and pair representations; a diffusion module predicts atomic coordinates and separate heads estimate confidence.SourcesAlphaFold 3 paper · Model architecture; Fig. 1d and Fig. 2 |
| Known versions | AlphaFold 3; inference code/documentation reviewed at commit c0f97eda2f1f482fd94d3a38bece18c7069b4a5c. This is a software revision, not a weight-file checksum. Paper evaluation variants are distinct configurations.Sources (2)AlphaFold 3 README; AlphaFold 3 paper · Pinned repository revision; paper evaluation distinctions described in Methods |
| Inputs | Protein, DNA and RNA sequences; chemical components; optional MSAs and structural templates.SourcesAlphaFold 3 input specification · Top-level structure; protein, RNA, DNA and ligand inputs |
| Outputs | Predicted structures in mmCIF plus confidence outputs, including pLDDT, PAE, pTM and ipTM.SourcesAlphaFold 3 output specification · Output directory structure; confidence outputs |
| Training cutoff | Experimental PDB structures plus protein and RNA distillation sets. The standard structural cutoff is 2021-09-30. The dedicated PoseBusters Methods section specifies a separate 2019-09-30 model; other training passages conflict with that date (see below).Sources (2)AlphaFold 3 paper; AlphaFold 3 supplementary information · Main paper Methods: Training regime and PoseBusters; Supplement Sections 2.5 and 5.2 |
| Training data | Experimental PDB structures, protein and RNA distillation sets, and transcription-factor examples used during fine-tuning. Sequence-search databases are separately versioned input resources.SourcesAlphaFold 3 supplementary information · Sections 2.2 and 2.5; Table 3; training-data discussion below |
| Context limits | The default largest compilation bucket is 5,120 tokens. The documentation supports larger inputs by configuration, subject to memory; this is not a universal architectural context limit.SourcesAlphaFold 3 performance documentation · Compilation buckets; predicting structures with more than 5,120 tokens |
| Access | Public inference implementation; weights downloaded directly from Google under separate non-commercial terms.SourcesAlphaFold 3 README · Obtaining Model Parameters; Installation and Usage |
| Code licence | Apache License 2.0.SourcesAlphaFold 3 code licence · LICENSE |
| Weights licence | Custom AlphaFold 3 Model Parameters Terms of Use, last modified 2024-11-09. Non-commercial use by or for non-commercial organisations; additional output and redistribution restrictions apply.SourcesAlphaFold 3 weights terms · Key things to know; Use restrictions |
| Parameters | No total trainable-parameter count is reported in the inspected main paper, supplementary architecture/training sections or implementation documentation. Layer dimensions do not establish a complete checkpoint total. · Not reported in inspected sourcesSources (2)AlphaFold 3 paper; AlphaFold 3 supplementary information · Main paper Model architecture; supplement Sections 3–5 and full-text parameter search; implementation documentation |
Evidence
Source checking verifies the cited claim or transcription. It does not establish independent reproduction.
Evidence table
Inspect claims, sources and review details
Trace each statement to its source and review. A context-only reference supports the record generally; it does not verify an individual field. Source checking does not reproduce an experiment.
One row per statement and cited source. Multiple citations are not independent evaluations. Shared locators are labelled explicitly.
1 evidence row matching the loaded filters
| Property and statement | Original source and location | Review and provenance |
|---|---|---|
| Relationship: family discovery-model-alphafold-3 Individual claims | ESMFold2 primary paper v1, Figure 2C Runs N’ Poses Figure 2C Runs N’ Poses, bar label AlphaFold3 (MSA); Appendix A.2.10 Version: 10.64898/2026.06.03.729735v1; posted 2026-06-04 | source checked automated source review · 2026-09-23 Audit detailsSource-backed evaluated identity only; no independent reproduction. Field: Claim: esmfold2-2026-runs-n-poses-method-alphafold3-msa-discovery-model-alphafold-3-identity-claim Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
Sources and history
View linked audit checks and correction history
Release 2026-09-23-2b89723c6dd9 · Record review: source checked
1 source records and release history
- ESMFold2 primary paper v1, Figure 2C Runs N’ Poses · Original source · 10.64898/2026.06.03.729735v1; posted 2026-06-04
Technical metadata and extraction receipts
Stable ID: esmfold2-2026-runs-n-poses-method-alphafold3-msa
- areas
- proteins-complexes
- source locator
- Figure 2C Runs N’ Poses, bar label AlphaFold3 (MSA); Appendix A.2.10
- missing metadata
- checkpoint revision: unreported; parameters: unextracted