Curiosity · AI Model
Google DeepMind AlphaFold 3
AlphaFold 3 is the third-generation structure-prediction model from Google DeepMind and Isomorphic Labs. Unlike AlphaFold 2, which specialised in single-protein folding, AlphaFold 3 predicts the joint structure of complexes — proteins with DNA, RNA, small molecules (ligands), ions, and antibodies — making it the most capable open scientific model for drug-discovery workflows.
Model specs
- Vendor
- Google DeepMind
- Family
- AlphaFold
- Released
- 2024-05
- Context window
- 1 tokens
- Modalities
- text
Strengths
- Predicts joint structures of proteins, DNA, RNA, small molecules, and ions
- Major accuracy lift over AlphaFold 2 on complexes
- Free research access via AlphaFold Server (daily quota)
- Open peer-reviewed Nature publication
Limitations
- Not an LLM — no token context; inputs are biological sequences and ligand SMILES
- Commercial use is restricted; Isomorphic Labs licenses for pharma
- Hallucinated structures are possible for under-represented complexes — validate experimentally
- Server is rate-limited; self-hosted inference is recently available via code release but compute-heavy
Use cases
- Predicting protein-ligand binding in early drug discovery
- Modelling nucleic-acid complexes (DNA/RNA)
- Antibody design and immunology research
- Academic structural biology studies
Benchmarks
| Benchmark | Score | As of |
|---|---|---|
| PoseBusters ligand accuracy | +50% vs prior SOTA | 2024 |
| Protein-protein interface LDDT | State-of-the-art | 2024 |
Frequently asked questions
What is AlphaFold 3?
AlphaFold 3 is a structure-prediction model from Google DeepMind and Isomorphic Labs. Unlike AlphaFold 2 (protein-only), AlphaFold 3 predicts the joint 3D structure of protein complexes with DNA, RNA, small-molecule ligands, ions, and antibodies.
Is AlphaFold 3 an LLM?
No — AlphaFold 3 is a diffusion-based scientific model, not a language model. Its inputs are biological sequences (FASTA) and ligand descriptions (SMILES), and its outputs are 3D atomic coordinates. We include it here because it is a landmark 'AI for science' model that teams often compare alongside LLMs.
Can I use AlphaFold 3?
Yes — DeepMind hosts a free AlphaFold Server for non-commercial research with a daily job quota. In late 2024 the source code was published for academic use. Commercial drug-discovery use requires a licence via Isomorphic Labs.
How accurate is AlphaFold 3?
On benchmarks like PoseBusters it roughly doubles the ligand-pose accuracy of prior best docking methods, and it matches or beats specialised methods on protein-protein interfaces. For novel chemistry or under-represented systems, experimental validation is still required.
Sources
- Google DeepMind — AlphaFold 3 — accessed 2026-04-20
- Nature paper — AlphaFold 3 — accessed 2026-04-20