Structure prediction

AlphaFold 3

(Work in Progress)

Structure prediction for proteins, DNA/RNA, ligands, ions, and modified residues. Supports co-folding.

Google DeepMind and Isomorphic Labs' AlphaFold 3 predicts the joint all-atom structure of biomolecular complexes: proteins, DNA, RNA, ligands (by CCD code, SMILES, or a user-provided CCD), ions, post-translational and nucleotide modifications, glycans, and covalent bonds. Here, MSAs come from the ColabFold MMseqs2 server by default, so the ~630 GB genetic databases are not needed; its own local data pipeline is available for an operator who has installed them.

Documentation ↗ Input parameters ↗ Examples ↗ Home page ↗ Paper ↗ Source code ↗ License ↗

No jobs before this; it will run immediately.

Not logged in. Register an account to keep track of your jobs and results, and add presets and bookmarks. Email results to this address when complete:

Ready

Example preset

Choose an example to fill in its settings and input structures.

Input mode

Input

Job name The input's name; a sanitised version of it names the output directory and files.

Molecules

Add one box per unique entity. Chain IDs are uppercase letters (A, or A, B for two copies); left blank, unused IDs are assigned. A ligand is a SMILES string, or one or more CCD codes (CCD_ATP; CCD_NAG, CCD_NAG, CCD_BMA for a multi-component ligand such as a glycan, joined with Bonded atom pairs). SMILES ligands cannot take part in bonds. An ion is its CCD code (MG, ZN, CA). Modifications take a CCD code at a 1-based position. MSA paths are optional precomputed A3M files on the runner; a protein given only one of the two runs with the other empty. Cyclic chains are not supported: AlphaFold 3 does not accept bonds within or between polymer chains.

Model seeds Comma-separated integer random seeds (modelSeeds); at least one is required. The model runs once per seed, and each run produces Diffusion samples structures.

Bonded atom pairs (JSON, optional) Native bondedAtomPairs: covalent bonds as pairs of [entity ID, 1-based residue index, atom name], e.g. [[["A", 12, "SG"], ["B", 1, "C25"]]] for a covalent ligand, or the bonds between the components of a multi-CCD ligand. Atom names come from the CCD; bonds within or between polymer chains are not supported.

User-provided CCD (mmCIF, optional) Native userCCD: chemical components in CCD mmCIF format, defining custom ligands or modified residues (referenced by component ID from a ligand's CCD codes or a modification), with named atoms for bonds and ideal coordinates for when RDKit cannot generate a conformer. Avoid underscores in custom component IDs.

AlphaFold 3 input (JSON)

A complete AlphaFold 3 input: one object in the alphafold3 dialect (name, modelSeeds, sequences of protein, rna, dna and ligand entities, optional bondedAtomPairs and userCCD, dialect "alphafold3", version 1-4), or an AlphaFold Server job list, which AlphaFold 3 converts. In this form: Bio Web writes this text to a file on the runner, so any unpairedMsaPath, pairedMsaPath, mmcifPath, or userCCDPath inside it must be an absolute path available there. Entities that already carry MSAs or templates keep them.

Input file (JSON) An AlphaFold 3 input JSON file, in the alphafold3 dialect or as an AlphaFold Server job list. In this form: Choose a file. Its contents are read here and sent as the input document.

MSAs and templates

MSAs and templates Where MSAs and templates come from for entities whose input does not already give them. ColabFold: each distinct protein sequence is sent to the MSA server below; in a complex of several distinct proteins the server's paired alignment is placed first in each chain's unpaired MSA, as the input docs recommend for custom pairing. RNA runs MSA-free and no templates are used. None: every chain runs from its sequence alone (private sequences, offline runs), at a large cost in accuracy. AlphaFold 3 data pipeline: Jackhmmer, Nhmmer and Hmmsearch over the ~630 GB genetic databases, which must be installed on the runner.

MSA server The ColabFold MMseqs2 API the ColabFold option sends protein sequences to.

Maximum template date YYYY-MM-DD (--max_template_date). The data pipeline ignores templates released after it, and only CCD components released before it may fall back to their model coordinates when RDKit cannot generate a conformer. The default is the AlphaFold 3 paper's date.

Deduplicate unpaired against paired MSAs --resolve_msa_overlaps, the paper's behaviour. Turned off automatically when the ColabFold option builds paired alignments, since deduplication would break their row pairing.

Inference

Diffusion samples Structures sampled per seed (--num_diffusion_samples). The default is 5.

Recycles Trunk recycling iterations (--num_recycles). The default is 10.

Number of seeds (optional) --num_seeds: run this many seeds, counting up from the input's single seed, which must then be the only one it gives. Blank uses the input's seeds as they are.

RDKit conformer iterations (optional) --conformer_max_iterations: raise RDKit's conformer search limit when a ligand fails with "Failed to construct RDKit reference structure". Blank keeps RDKit's default.

Keep leaving atoms of standalone glycans --fix_standalone_glycans. AlphaFold 3 was trained with leaving atoms removed even from glycans bonded to nothing; this keeps them, away from the regime it was trained in.

Run inference Off: stop after the MSA and template step and return the completed input JSON (with the ColabFold or data-pipeline MSAs in it), for reuse in a later run.

Output

Save embeddings Write each seed's final single and pair embeddings (--save_embeddings); these are large float16 arrays.

Save distogram Write each seed's final distogram (--save_distogram), a num_tokens x num_tokens x 64 float16 array.

Compress large output files Write the mmCIF and confidences JSON zstandard-compressed (--compress_large_output_files). Compressed structures are not shown in the viewer.

Runtime

Attention implementation --flash_attention_implementation. Choose XLA for a compute capability 7.x GPU (V100, T4, RTX 20xx); the runner then also sets the XLA flag AlphaFold 3 requires for those devices.

JAX backend --jax_backend. CPU inference needs the XLA attention implementation.

GPU device Zero-based JAX device index (--gpu_device), after any CUDA_VISIBLE_DEVICES filtering.

Unified memory Let GPU memory spill into host memory instead of failing out of memory, for GPUs with less than 80 GB: sets XLA_PYTHON_CLIENT_PREALLOCATE=false, TF_FORCE_UNIFIED_MEMORY=true and XLA_CLIENT_MEM_FRACTION=3.2, as the performance docs recommend. Slower.

Compilation buckets (optional) --buckets: strictly increasing, comma-separated token counts that inputs are padded to, so a compiled model can be reused. Blank keeps AlphaFold 3's defaults (256 to 5120).

Ready

Endpoint POST /api/alphafold3/

Send JSON or standard form data. A GET request returns this tool's field schema.

API reference for AlphaFold 3 →

License

Non-commercial. The source code is Apache 2.0, but the model parameters and every prediction made with them are covered by the AlphaFold 3 Model Parameters Terms of Use: only for non-commercial use by, or on behalf of, non-commercial organisations (universities, non-profits, research institutes, education, journalism, government); never in connection with commercial activities, including research on behalf of commercial organisations; outputs may not be used to train structure-prediction models; and the parameters may not be shared outside your organisation. Outputs are subject to the AlphaFold 3 Output Terms of Use.