Boltz-2

Boltz structure and property prediction, wrapped for subseq.bio.

Inputs and outputs

Your files are mounted read-only under /inputs inside the container.
Prediction outputs and metadata should be written to /outputs.
SubSeq manages the model cache automatically.
The container entrypoint is boltz predict, so the job arguments are passed directly to that CLI.
If your YAML does not include MSAs, set msa: empty for each protein chain, or Boltz-2 may fail expecting MSA input.
To use a precomputed MSA, include the alignment file under /inputs and reference it from YAML, for example msa: /inputs/target.a3m. For multi-protein custom MSAs, use Boltz's paired CSV format.
Boltz-2 affinity is small-molecule-only. Guided affinity rejects SMILES ligands above 128 non-hydrogen atoms before scheduling; use a polymer chain for protein, DNA, RNA, or peptide binders.
Use polymer chain entries for protein, DNA, RNA, or peptide binders. Do not encode polymers as SMILES; use residue modifications or custom YAML for modified residues.
When using directory input, point Boltz-2 at a directory containing only .yaml/.fasta inputs; keep auxiliary files such as .a3m outside that scanned directory and reference them by absolute /inputs/... paths.

Guided modes

Single Sequence: generate YAML for one protein chain.
Build Assembly: generate YAML from a repeatable protein/DNA/RNA chain list, optional per-chain CCD residue modifications/cyclic flags, and optional small molecule. For dsDNA or paired RNA, add two nucleic-acid chains and enter the second strand as the reverse complement; the guided form can add that chain from an existing DNA or RNA chain. Turn on Predict Affinity to add properties: affinity for the small-molecule ligand. Guided affinity is limited to protein-only polymer targets.
Custom Input: run one supplied Boltz YAML or FASTA file. Use this for paired MSAs, templates, constraints, large/symmetric assemblies, and other advanced schemas.
Batch Folder: run all Boltz YAML/FASTA inputs in a folder as one SubSeq job.

Affinity metrics are produced only when the YAML contains properties: affinity. In guided mode that means Build Assembly with a small molecule and Predict Affinity enabled. In Custom Input or Batch Folder, each YAML file can produce affinity metrics if it contains a valid affinity property.

Example 1 — basic run

Use the default pattern shown on the New Job form: point Boltz-2 at your inputs directory and turn on potentials.

/inputs
--use_potentials
--out_dir=/outputs

Upstream recommends YAML inputs; FASTA inputs are supported but deprecated upstream.
Place your YAML/FASTA and any required assets under /inputs when uploading.
--use_potentials enables the learned potentials during inference.
--out_dir=/outputs directs all prediction artifacts to the job output volume.

Example 2 — local configuration file

You can also drive Boltz-2 with an explicit configuration file that lives under /inputs.

/inputs/my_config.yaml
--out_dir=/outputs/my_config_run
--use_potentials

/inputs/my_config.yaml is a Boltz configuration file you prepare based on the upstream examples.
Outputs for this run will be written under /outputs/my_config_run.
The platform manages the model cache automatically.

Minimal no-MSA YAML pattern:

sequences:
  - protein:
      id: A
      sequence: "MKTAYIAKQRQISFVKSHFSRQDILDLI"
      msa: empty

Modified residues can be entered in guided Build Assembly mode, or specified directly in custom YAML:

version: 1
sequences:
  - protein:
      id: A
      sequence: "MSTNPKPQRKTKRNTNRRPQDVKFPGG"
      msa: empty
      modifications:
        - position: 2
          ccd: MSE

Example 3 — multimer from YAML

Guided Build Assembly mode can generate multiple protein, DNA, or RNA chains. Protein chains use msa: empty. For full control, create a YAML input under /inputs/multimer.yaml with multiple polymer chains:

sequences:
  - protein:
      id: A
      sequence: "MKTAYIAKQRQISFVKSHFSRQDILDLI"
      msa: empty
  - dna:
      id: B
      sequence: "ATCGATCG"

A DNA duplex is specified as two DNA chains, with both strands entered in their own 5-prime to 3-prime direction. In guided mode, use Add Reverse Complement after entering a DNA or RNA chain to append the reverse-complement strand.

version: 1
sequences:
  - dna:
      id: A
      sequence: "ATGCCGTA"
  - dna:
      id: B
      sequence: "TACGGCAT"

Then run:

/inputs/multimer.yaml
--use_potentials
--out_dir=/outputs/multimer

For de novo binder screening, a common pattern is to keep the designed binder chain as msa: empty, while giving the natural target chain a real MSA when available.

Quality-oriented sampling

The upstream defaults are already the normal full prediction path: --sampling_steps=200, --recycling_steps=3, --diffusion_samples=1, --step_scale=1.5, and --max_msa_seqs=8192.

/inputs/my_complex.yaml
--out_dir=/outputs/my_complex_seed01
--use_potentials
--recycling_steps=10
--sampling_steps=200
--diffusion_samples=1
--max_parallel_samples=1
--step_scale=1.5
--max_msa_seqs=8192
--output_format=pdb
--write_full_pae
--write_full_pde
--seed=12345
--override

Boltz documents --recycling_steps=10 --diffusion_samples=25 as an AlphaFold3-like high-cost sampling setting.
For comparing sampled complex poses on SubSeq, separate one-sample jobs with different --seed values are often easier to inspect and retry than one bundled high-sample job.
For unbiased binder screening, avoid hotspot/interface constraints unless you intentionally want constrained prediction; constraints can hide interface failure.

Additional CLI options

SubSeq passes arguments directly to boltz predict, while managing the model cache automatically.

Show common boltz predict options

--out_dir PATH
--recycling_steps INT
--sampling_steps INT
--diffusion_samples INT
--max_parallel_samples INT
--step_scale FLOAT
--output_format [pdb|mmcif]
--num_workers INT
--preprocessing-threads INT
--override
--seed INT
--max_msa_seqs INT
--subsample_msa
--num_subsampled_msa INT
--no_kernels
--use_potentials
--method STRING
--model [boltz1|boltz2]
--affinity_mw_correction
--sampling_steps_affinity INT
--diffusion_samples_affinity INT
--write_full_pae
--write_full_pde
--write_embeddings

For detailed defaults and the YAML schema, see the upstream Boltz prediction instructions.

Notes

Confidence files such as confidence_<id>_model_0.json include confidence_score, ptm, iptm, complex_plddt, complex_iplddt, complex_ipde, chains_ptm, and pair_chains_iptm.
Custom A3M files should be plain text. If an upstream MSA tool writes a trailing NUL byte, strip it before passing the A3M to Boltz-2.
For full CLI options and configuration schemas, see the Boltz GitHub repository (boltz-2 section).
To start a job from the UI, go to New Job → Boltz-2 and paste one of the argument blocks above.