Structure prediction

Protenix-v2

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

Protenix is ByteDance's trainable, open reproduction of AlphaFold 3. It predicts all-atom structures of biomolecular complexes, including proteins, DNA, RNA, ligands, ions, and covalent modifications, with optional MSA, template, RNA MSA, and pocket or contact constraint inputs.

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 Names the single Protenix job generated from these parameters; results are written under this name.

Molecules

Add one box per unique Protenix entity. Set count and an optional matching list of comma-separated chain IDs, plus the sequence, ligand (a CCD_ code such as CCD_ATP, several joined for a glycan such as CCD_NAG_BMA_BGC, a SMILES string, or FILE_ and an absolute path to a 3D structure file), or ion code (e.g. MG, without CCD_). Modifications take a CCD code at a 1-based position. "Cyclic" closes a peptide with a head-to-tail amide bond. MSA and template paths are optional precomputed files.

Covalent bonds (JSON, optional) Native Protenix covalent_bonds array, e.g. [{"entity1": "2", "copy1": 1, "position1": "2", "atom1": "N6", "entity2": "3", "copy2": 1, "position2": "1", "atom2": "C1"}]. Entities are 1-based box indexes; copies are 1-based within count and may be omitted together to bond every copy pair.

Constraint (JSON, optional) Native Protenix constraint object with a "contact" list and/or a "pocket" object, e.g. {"pocket": {"binder_chain": {"entity": 2, "copy": 1}, "contact_residues": [{"entity": 1, "copy": 1, "position": 69}], "max_distance": 8}}. A soft constraint, read only by protenix_base_constraint_v0.5.0.

Protenix inference jobs (JSON) Complete native Protenix input: a non-empty top-level list of named jobs using proteinChain, dnaSequence, rnaSequence, ligand, or ion entities, with optional covalent_bonds, constraint, and modelSeeds. In this form: Bio Web writes this text to input.json. Any MSA, template, or FILE_ ligand paths inside it must be absolute paths available to the runner.

Inference jobs file (JSON) The JSON file defining this run's jobs, in Protenix's own input format. In this form: Choose a file. Its contents are read here and sent as the input document.

Inference

Model Checkpoint to run; weights download on first use. protenix-v2 and the v1.0.0 base models add template and RNA MSA support; only protenix_base_constraint_v0.5.0 reads constraints; the mini and tiny models trade accuracy for speed (recommended: 4 cycles and 5 steps), and the ESM and ISM variants add ESM2-3B embeddings for use without MSAs. The protenix pred CLI default is protenix_base_default_v1.0.0.

Seeds Comma-separated integers passed to --seeds; every job runs once per seed. The CLI default is 101.

Use seeds from the input JSON Use modelSeeds from the first job in place of Seeds, for every job in the run.

Samples Number of structure samples per seed (--sample). The CLI default is 5.

Pairformer cycles Number of Pairformer recycling cycles (--cycle). The CLI default is 10.

Diffusion steps Number of diffusion steps (--step). The CLI default is 200.

Use the model's recommended parameters Replace cycles and steps with the selected model's recommended values: 10 and 200 for base models and protenix-v2, 4 and 5 for mini and tiny models. Also turns MSAs off for the mini ESM and ISM models.

Optional features

Use or generate protein MSAs Use protein MSA features. A protein without paired or unpaired MSA paths has its sequence sent to the MSA server chosen below. Turn off for private sequences or offline runs, at a significant cost in accuracy unless an ESM model is used.

MSA server Where missing protein MSAs are searched: the Protenix MSA service (https://protenix-server.com/api/msa; set MMSEQS_SERVICE_HOST_URL on the runner to use your own) or the public ColabFold MMseqs2 server (https://api.colabfold.com).

Use protein templates Use template features from each protein's templatesPath (.a3m or .hhr). A protein without one gets a local template search instead, which needs protein MSAs, HMMER (hmmsearch and hmmbuild) on the runner's PATH, and downloads the PDB seqres database on first use. Requires protenix-v2 or a v1.0.0 base model.

Use RNA MSAs Use RNA MSA features from each RNA's unpaired MSA path. An RNA without one gets a local nhmmer search instead, which needs HMMER on the runner's PATH and downloads the Rfam, RNAcentral, and NT-RNA databases on first use. Requires protein MSAs to be on, and protenix-v2 or a v1.0.0 base model.

Use training-free guidance Apply Training-Free Guidance (--use_tfg_guidance) for improved ligand plausibility, such as chirality and planarity; increases compute time.

Output

Write atom-level confidence data Write full per-atom confidence JSON in addition to each sample's summary confidence.

Runtime

Inference precision Protenix defaults to BF16; FP32 uses more memory.

Triangle attention kernel Kernel for triangle attention. DeepSpeed needs DeepSpeed and CUTLASS_PATH on the runner; PyTorch needs no extra kernels.

Triangle multiplicative kernel Kernel for the triangle multiplicative update.

Cache diffusion variables Cache shareable variables in the diffusion module.

Kernel fusion Enable efficient kernel fusion in the diffusion transformer.

TF32 matrix multiplication Enable TF32 for FP32 matrix multiplications on supported GPUs.

Ready