AI model developed by Google DeepMind and Isomorphic Labs for predicting 3D structures and interactions of biomolecules including proteins, DNA, RNA, ligands, ions, and modified residues.
Parameters
Nieujawnione publicznie
parameters
Release date
8 May 2024
Access:HostedDownloadDeployment:☁ Cloud💻 Local
Overview
Applications
Access & deployment
HostedDownload
CloudLocal
Weights: Open weights
Key parameters
🧩 Parameters: Nieujawnione publicznie
📥 Input: structured data
Technical specification
Parameters
Nieujawnione publicznie
parameters
License
Kod źródłowy: CC BY-NC-SA 4.0; wagi modelu: AlphaFold 3 Model Parameters Terms of Use (wyłącznie niekomercyjne, akademickie – na wniosek)
Hardware requirements
Minimum: GPU with 16 GB VRAM (sequence length limit of 1280 tokens); recommended: NVIDIA A100 40 GB or H100. Accuracy benchmarks were conducted exclusively on A100 and H100.
Modalities
⬇ Input
structured_data
⬆ Output
3d_representationresearch_outputsstructured_data
Capabilities and applications
Native model capabilities
Reasoning
The model's ability to reason logically and solve complex problems.
Category: reasoning
Structured output
Producing data in structured formats such as JSON.
Category: structured_generation
Benchmark results
8 benchmarks
DockQ
76.6%
PoseBusters
93.2%
PoseBusters V1 – protein-ligand docking (blind)
% PB-valid poses z ligand RMSD < 2 Å · PoseBusters benchmark (428 protein-ligand structures from PDB ≤ 2021), blind mode (no binding pocket information provided). Confirmed by EBI: Chai-1 achieves 77% (vs. 76% for AF3).
76%
📅 8 May 2024📄 Abramson et al., Nature 630, 493–500 (2024); EBI AlphaFold training page
AlphaFold 3 is the first AI system to surpass physicochemical docking tools on this benchmark. The score improves when binding pocket information is provided.
PoseBusters – poprawa vs. najlepsze tradycyjne metody
Względna poprawa dokładności (PB-valid) względem najlepszej tradycyjnej metody (blind docking) · Comparison against Vina and GOLD on the PoseBusters V1 benchmark set; AlphaFold 3 run without structural information about the binding pocket.
+50%
📅 8 May 2024📄 Isomorphic Labs blog (May 2024); Wikipedia AlphaFold
The claim originates from the official announcement by Isomorphic Labs and Google DeepMind. Independent analyses suggest that stronger docking tools may offset part of this advantage.
CASP15 RNA – protein-nucleic acid accuracy
Interface LDDT; % struktur z RMSD < 2 Å · CASP15 RNA dataset; AF3 compared against RoseTTAFold2NA and AIchemy_RNA. AIchemy_RNA2 (with human intervention) performed marginally better.
wyższy niż RoseTTAFold2NA%
📅 8 May 2024📄 Abramson et al., Nature 630, 493–500 (2024)
AlphaFold 3 outperforms RoseTTAFold2NA on protein–nucleic acid complexes; for RNA monomers from CASP15, AIchemy_RNA2 with expert assistance was marginally better.
Protein monomer LDDT (Recent PDB eval set)
LDDT · Evaluation of protein monomers on a PDB set with entries newer than the training cutoff date (30.09.2021).
wyższy niż AlphaFold-Multimer v2.3
📅 8 May 2024📄 Abramson et al., Nature 630, 493–500 (2024)
Local improvement over AlphaFold 2; global improvement is limited according to independent benchmarking (PMC12661943).
Antibody-antigen prediction accuracy vs. AlphaFold-Multimer v2.3
DockQ / % poprawnych interfejsów · Recent PDB evaluation set; protein–antibody interfaces.
znacząco wyższy%
📅 8 May 2024📄 Abramson et al., Nature 630, 493–500 (2024)
Statistically significant improvement (p < 0.001) over AlphaFold-Multimer v2.3.
Covalent modifications – bonded ligands success rate
% struktur z RMSD < 2 Å · Covalently modified protein residues and ligands; accuracy range of 40–80% depending on category.
~80%
📅 8 May 2024📄 Abramson et al., Nature 630, 493–500 (2024)
~40% for modified RNA residues, ~80% for covalent ligands.
Deployment and security
🔒 Security / Enterprise
✓ Verified enterprise information
No classic enterprise/SaaS-style public security package is available. Google DeepMind describes a responsible release process for AlphaFold 3, including consultations with biosecurity experts and restrictions on non-commercial use.
Security information primarily concerns responsible release and usage restrictions, not enterprise compliance features such as SSO, SOC 2, or data residency.
Updated: 15 Mar 2026↗ Security documentation
Sources and related pages
9 sources
LinkAlphaFold — Google DeepMindLinkAlphaFold 3 predicts the structure and interactions of all of life’s moleculesLinkgoogle-deepmind - alphafold3LinkAlphaFold ServerLinkAlphaFold 3 Model Parameters Terms of UsePaperAccurate structure prediction of biomolecular interactions with AlphaFold 3 – Nature 630, 493–500 (2024)BlogAlphaFold 3 predicts the structure and interactions of all of life's molecules – Isomorphic LabsDocsHow does AlphaFold 3 work? – EBI AlphaFold trainingPaperA comprehensive benchmarking of AlphaFold 3 – PMC12661943