Boltz-2 (MolComp)

Quick links

• Paper: https://www.biorxiv.org/content/10.1101/2025.06.14.659707v1
• GitHub: https://github.com/jwohlwend/boltz

What MolComp runs

MolComp runs the upstream boltz predict CLI inside an AWS Batch container. Your Playground inputs are translated into:

• an input.yaml (either pasted YAML or extracted from your uploaded bundle)
• optional CLI flags (from the Playground parameter form)

All outputs produced under the run output directory are uploaded to your Workspace:
user/<your_id>/jobs/<job_id>/

Inputs

Option A: Paste YAML (best for simple inputs)

Playground field: “Boltz input YAML”

If your YAML references extra files (MSAs, templates, ligand files, etc.), you must instead use the bundle option below, otherwise the run will fail because those files will not exist in the container.

Option B: Upload a bundle zip (recommended for anything non-trivial)

Playground field: “Input bundle (.zip)”

Zip requirements:

• Must contain input.yaml (or input.yml)
• Any referenced paths in YAML (MSA files, templates, etc.) should be included in the zip at the same relative paths
• The runner extracts the zip into the job working directory and runs boltz predict input.yaml ...

Bundle precedence:

• If you upload a bundle, it takes precedence over pasted YAML.

Parameters (how MolComp maps them)

MolComp uses two mechanisms:

1) Structured fields in the Playground (the common parameters)
2) “Advanced (JSON merge)” which merges arbitrary key/value pairs into params

Runner behavior:

• Any params.<key> (except extra_params) is passed through to the CLI as:
  • --<key> <value> for numbers/strings
  • --<key> for boolean true
  • omitted for boolean false / empty values

This means you can use the Advanced JSON block to access any upstream CLI flag immediately, without waiting for a UI update.

Common parameters in the Playground

These correspond to the most frequently used inference knobs:

• recycling_steps (int)
• sampling_steps (int)
• diffusion_samples (int)
• max_parallel_samples (int)
• step_scale (float)
• use_potentials (bool)
• use_msa_server (bool) + msa_server_url (string)
• msa_pairing_strategy (string)
• output_format (string: mmcif or pdb)
• num_workers (int)
• write_full_pae, write_full_pde, write_full_plddt (bool)

Advanced JSON merge examples

Example: enable full matrices and tune diversity

{
  "write_full_pae": true,
  "write_full_pde": true,
  "write_full_plddt": true,
  "step_scale": 2.0,
  "diffusion_samples": 5
}

Example: set runtime/execution knobs (only if your upstream version supports them)

{
  "devices": 1,
  "accelerator": "gpu",
  "seed": 7
}

Outputs

MolComp uploads the full contents of the Boltz output directory to Workspace.

Commonly you will see:

• predicted structure files (PDB/mmCIF)
• confidence metrics (JSON/NPZ)
• affinity outputs (if requested in YAML)

Exact filenames can vary by upstream version; MolComp intentionally uploads the entire output folder so you don’t lose anything.

Notes / gotchas

• YAML that references files requires a bundle zip.
• Large MSAs and templates can substantially increase runtime; adjust est_gpu_hours accordingly.
• If you need a CLI flag that isn’t exposed as a first-class UI control, use “Advanced (JSON merge)”.