Pub-AI logoPub-AI: AI in Science
domain-discoveryfreemiumbiologychemistryhealth-medicineprotein-design
Domain DiscoveryFreemium🔗 source linked🔓 open source

AlphaFold 3

AlphaFold 3 extends AlphaFold 2's protein structure prediction to handle DNA, RNA, small molecules, ions, and modified residues simultaneously — enabling protein-ligand, protein-DNA, and protein-RNA complex prediction within a single diffusion-based model.

Pricing noteAlphaFold Server is free for non-commercial academic use (up to 20 seed sequences per job). Model weights are available under a non-commercial research license; commercial use requires a license from Isomorphic Labs.
Last verified: September 2026

What it does

AlphaFold 3 (May 2024) is a structure prediction model from Google DeepMind and Isomorphic Labs that can predict the 3D structure of complexes involving proteins, DNA, RNA, small molecules (ligands), and ions in a single unified model.

The key advance over AlphaFold 2: AF2 was limited to protein chains. AF3 accepts arbitrary combinations of biomolecular inputs and predicts how they fold and interact with each other. This makes it directly applicable to:

  • Drug discovery — protein-ligand docking and binding pose prediction
  • Structural genomics — protein-DNA and protein-RNA complexes
  • Antibody design — antibody-antigen complex prediction
  • Enzyme engineering — substrate binding and active-site geometry

The underlying architecture is also fundamentally different: AF2 used an Evoformer transformer trained with end-to-end supervised learning. AF3 uses a diffusion model (similar in principle to image generation models like Stable Diffusion) trained to denoise 3D atomic coordinates from noise. This makes predictions probabilistic — AF3 generates a distribution of structures and samples from it.

Using the AlphaFold Server

The AlphaFold Server at alphafoldserver.com provides free access for non-commercial research:

  1. Enter protein sequences (FASTA format), nucleic acid sequences, or SMILES strings for small molecules
  2. Specify how many copies of each component are in your complex
  3. Submit — jobs typically complete in 10–30 minutes depending on queue
  4. Download the predicted structure as a CIF file, along with confidence metrics

Inputs the server accepts:

  • Protein sequences (standard amino acids plus some modified residues)
  • DNA sequences (single-stranded or double-stranded)
  • RNA sequences
  • Small molecules via SMILES notation
  • Ions

What you get back:

  • Predicted 3D structure in CIF format (viewable in UCSF ChimeraX, PyMOL, Mol*, etc.)
  • pLDDT scores per residue (confidence, same scale as AF2)
  • PAE (predicted aligned error) matrix for the complex
  • ipTM score for the interface (key metric for assessing binding confidence)

How it compares to AlphaFold 2

AlphaFold 2 remains the reference tool for single-protein structure prediction and outperforms AF3 on some single-chain benchmarks — the diffusion architecture trades some single-chain accuracy for the ability to handle multi-component systems.

For drug discovery and ligand binding, AF3 is substantially better than AF2 because it can predict where a small molecule sits within a protein binding pocket. AF2 cannot predict ligand poses at all without modifications.

For pure protein structure prediction where no ligand or nucleic acid is involved, the accuracy difference is small enough that AF2 (still freely available via the ColabFold notebook) remains widely used.

Limitations

Diffusion sampling introduces variability. Unlike AF2 which gives a single deterministic prediction, AF3 samples structures from a distribution. Running the same job twice can give slightly different results. High pLDDT and ipTM scores correlate with consistent, confident predictions; low scores indicate regions of genuine structural uncertainty.

Ligand predictions require validation. Protein-ligand complex predictions from AF3 have shown promising but imperfect accuracy on benchmark datasets. Before treating a predicted binding pose as ground truth, confirm with orthogonal methods (docking software like AutoDock Vina, or ideally experimental crystallography).

No dynamics. Like AF2, AF3 predicts a static structure. It does not predict conformational changes, allosteric effects, or the timescales of binding and release. For those questions, molecular dynamics simulation remains necessary.

Non-commercial license on weights. The model weights available for local download are non-commercial. Academic groups can run AF3 on their own infrastructure, but commercial use requires contacting Isomorphic Labs.

  • AlphaFold (2) — the original, still widely used for single protein chains
  • ESMFold — Meta’s protein language model–based predictor; faster, no multiple sequence alignment needed
  • RFdiffusion + ProteinMPNN — protein design (creating new sequences), not just structure prediction
← Back to Tools Directory