Protein Design Boltzgen

Protein Design (BoltzGen)

Overview

This skill runs the FastFold BoltzGen workflow in the same flow used by the Composer:

1. Prepare or generate design-spec YAMLs (single-spec or multi-spec).
2. Fetch/clean CIF inputs or reuse existing references.
3. Create a draft workflow.
4. Upload files and upsert the full workflow graph (workflow.yml) in one transaction.
5. Share the full Composer link and a draft review summary (inputs + params + YAML previews) after non-empty graph upsert.
6. Ask user to validate and confirm.
7. Execute, wait, and summarize candidate metrics + structure links.

It is based on:

• BoltzGen upstream repository and README: https://github.com/HannesStark/boltzgen
• BoltzGen paper (bioRxiv DOI 10.1101/2025.11.20.689494): https://www.biorxiv.org/content/10.1101/2025.11.20.689494v1.full

Base URLs

Keep script defaults aligned with other skills:

• API default: https://api.fastfold.ai
• UI default for user-facing links: https://cloud.fastfold.ai

Authentication

Preferred (Sandwalk agent): Integrations → Fastfold AI Cloud, exposed as FASTFOLD_API_KEY in the chat shell. This skill sends X-API-Key: <FASTFOLD_API_KEY>. Use env vars Sandwalk sets — do not hardcode or search home config paths.

Do not ask users to paste secrets in chat. Do not create .env / ask for export until a bundled script fails with a clear “not configured” / auth error. Prefer printenv FASTFOLD_API_KEY or run the script.

Env vars: FASTFOLD_API_KEY, SANDWALK_SESSION_WORKSPACE (workspace root), SANDWALK_SKILLS_DIR (skills root when set).

Alternatives: workspace .env (references/.env.example), shell export, or create a key at https://cloud.fastfold.ai/api-keys.

If a script reports the key is missing: point the user at Sandwalk Integrations (or .env / export), then retry. Do not hunt the filesystem for secrets.

When to Use This Skill

• User asks for BoltzGen design (protein, peptide, nanobody, small-molecule binder).
• User needs design-spec authoring help (single or multiple YAML specs).
• User needs CIF/PDB fetch/cleanup before upload.
• User wants an API-only workflow equivalent to Composer.
• User asks for draft workflow/composer links and run status.
• User asks to interpret BoltzGen output metrics and candidates.

Scripts

This skill bundles self-contained scripts under scripts/ (stdlib only).

Sandwalk agent — how to invoke (first match wins):

1. If SANDWALK_SKILLS_DIR is set: python3 "$SANDWALK_SKILLS_DIR/protein_design_boltzgen/scripts/<name>.py ...".
2. Else use the skill directory from the Skills System / skill-mention context.
3. Put downloads under "$SANDWALK_SESSION_WORKSPACE" when set.

Do not treat /protein_design_boltzgen, /skills/..., or Modal /workspace as the install path. Do not hardcode ~/.sandwalk/... or hunt with find/locate.

Primary scripts:

• python scripts/workflow_api.py — workflow create/build/upload/execute/results
• python scripts/fetch_cif.py — fetch input CIF files

Critical execution guardrail (non-negotiable)

If python scripts/workflow_api.py or python scripts/fetch_cif.py returns an error:

1. Report the exact command + concise error.
2. If FASTFOLD_API_KEY is unset, point the user at Sandwalk Integrations (or .env / export; create a key at https://cloud.fastfold.ai/api-keys).
3. Stop. Do not attempt fallback discovery (find, locate, ls package trees, python -c).

Background execution protocol (required)

When users ask to run BoltzGen "in background", use this split:

1. Run draft/submit/execute in foreground.
2. Capture and print workflow_id immediately.
3. Only background the long wait/log watch step.
4. Fetch results using the same preserved workflow_id.

Non-negotiable rules:

• Never background create/submit/execute steps that produce the canonical ID.
• Never ask the user to recover workflow_id for an agent-initiated run.
• Never use filesystem hunting for ID recovery (find, locate, ls /tmp, shell history grep).
• If ID capture failed due command error, rerun submit in foreground and return the new workflow_id.

Fast path for "show examples"

For prompts like "Show me Boltzgen protein design examples":

1. Run python scripts/workflow_api.py example-files --list.
2. Present that output directly.
3. Optionally run one preset resolution command (--preset ... --json) if user asks for details.
4. Do not scan directories unless the user explicitly requests file-level inspection.

• Create draft workflow:
  • explicit name:
    • python scripts/workflow_api.py new --name "API - vanilla target binding site"
  • auto simple name:
    • python scripts/workflow_api.py new --preset vanilla_target_binding_site
• Resolve bundled example files (recommended first step for examples):
  • list available presets:
    • python scripts/workflow_api.py example-files --list
  • resolve files for a preset:
    • python scripts/workflow_api.py example-files --preset vanilla_target_binding_site --json
  • quick alias for 5CQG:
    • python scripts/workflow_api.py example-files --preset 5cqg --json
• Build workflow spec from official template (after uploads):
  • python scripts/workflow_api.py build-spec --preset vanilla_target_binding_site --out /tmp/boltzgen_workflow.yml
  • python scripts/workflow_api.py build-spec --preset 5cqg --json
• Fetch and clean CIF:
  • python scripts/fetch_cif.py --pdb-id 5cqg --out-dir /tmp/boltzgen_inputs
• Upload file to workflow workspace:
  • python scripts/workflow_api.py upload --file /tmp/boltzgen_inputs/5cqg.cif --file-type protein
  • python scripts/workflow_api.py upload --file ./my_design.yaml --file-type yml
• Save graph spec (single upsert):
  • python scripts/workflow_api.py upsert --spec /tmp/workflow.yml
• Draft review for user validation (after upsert):
  • python scripts/workflow_api.py draft-review
  • python scripts/workflow_api.py draft-review --json
  • includes:
    • upserted workflow.yml preview
    • per design-spec YAML preview (binding-site fields visible before run)
• Print/share full Composer link (after upsert):
  • python scripts/workflow_api.py composer-link
• Execute:
  • python scripts/workflow_api.py execute
• Wait (auto-downloads all artifacts on COMPLETED):
  • python scripts/workflow_api.py wait --poll-seconds 30 --timeout-seconds 7200
  • choose output dir: python scripts/workflow_api.py wait --out-dir ./fastfold-artifacts/protein_design_boltzgen/<workflow_id>
  • opt out of auto-download: python scripts/workflow_api.py wait --no-download
• Logs (single snapshot + interpretation):
  • python scripts/workflow_api.py logs
  • python scripts/workflow_api.py logs --tail-lines 200
• Live logs while running:
  • python scripts/workflow_api.py logs --watch --poll-seconds 30 --timeout-seconds 1800
• Logs JSON payload:
  • python scripts/workflow_api.py logs --json
• Get candidates/metrics + links:
  • python scripts/workflow_api.py results
  • python scripts/workflow_api.py results --json (includes full parsed_results_raw, all metric field names, and ranked_table)
• Download all output artifacts (CIF/CSV/PDF) to a local directory (writes INVENTORY.md + inventory.json):
  • python scripts/workflow_api.py download
  • python scripts/workflow_api.py download --out-dir ./fastfold-artifacts/protein_design_boltzgen/<workflow_id>
  • python scripts/workflow_api.py download --json
  • default out dir: ./fastfold-artifacts/protein_design_boltzgen/<workflow_id>; per-file cap via --max-bytes (default 200 MB)

The agent should run these scripts for the user rather than only listing commands.

Downloading results & artifacts

Results return output files as library item references (libraryItemId + fileName), not direct URLs. Always download with the bundled script — never claim files are browser-only, and never use the browser route cloud.fastfold.ai/api/structure (it rejects API keys). The correct, API-key path is GET /v1/library/file/{item_id}/{file_name} (note the library/file/... order); the script resolves the signed artifacts.fastfold.ai URL and saves the bytes.

• When a run reaches COMPLETED, download all artifacts automatically — do not wait to be asked:
  • wait auto-downloads on COMPLETED (pass --no-download only if the user opts out).
  • If the run already finished, run python scripts/workflow_api.py download yourself.
• For non-COMPLETED terminal states (FAILED/STOPPED/CANCELLED) there are no artifacts — explain the status instead.
• Each download writes a manifest into the output dir: INVENTORY.md (readable table: file, type, size, description, CSV column preview) and inventory.json (full column list). Point the user to INVENTORY.md.

Artifact types: rank<k>_<x>.cif (ranked design #k structure), all_designs_metrics.csv (all designs), final_designs_metrics_<budget>.csv (final selected set), results_overview.pdf (report).

After downloading, report the local out_dir paths plus the Composer/Mol* links. See references/api_endpoints.md.

Citing local files: After download / wait auto-download, copy every path from local_paths or DOWNLOADED_LOCAL_PATHS: verbatim — full absolute paths, one per line. Never middle-truncate filenames or workflow ids. For remote URLs, print Label: https://… (bare URL visible), not markdown [Label](url).

Example-First Behavior (Required)

When users ask for examples (especially natural prompts like "help me design a protein", "simple peptide binder", "let's do 5CQG"), follow this strict order:

1. Use bundled local preset files first from references/examples/.
2. Do not fetch from PDB/RCSB if a matching local preset file exists.
3. Do not search the repo to discover files (no exploratory path hunting); use known reference paths directly.
4. Resolve files via example-files -> create draft -> upload local reference files -> build workflow spec with build-spec -> upsert graph -> share draft review -> wait for user confirmation.

Default mapping for natural requests:

• Generic "simple peptide binder" -> vanilla_target_binding_site
• "5CQG" -> references/examples/vanilla_target_binding_site/beetletert.yaml
  • references/examples/vanilla_target_binding_site/5cqg.cif
• "simple protein example" -> vanilla_protein

Only fetch/clean CIF from external sources when:

• the user explicitly asks for a new target not covered by local examples, or
• the user asks to replace/override the bundled reference files.

For preset smoke tests, do not hand-write workflow YAML. Always use:

• references/workflow_specs/*.workflow.yml (official templates)
• python scripts/workflow_api.py build-spec ... (placeholder replacement)

From-Scratch YAML Authoring (When no preset fits)

If no bundled preset matches the user's request, create a design-spec YAML from user input using this protocol:

1. Collect required inputs from the user first:
   • target structure file (.cif preferred; .pdb allowed)
   • target chain(s) and optional residue ranges
   • design modality (protein / peptide / nanobody / ligand context)
   • design constraints (binding site residues, include/exclude, insertions, bonds)
2. Ask user to provide/upload the target CIF/PDB file before drafting final YAML.
3. Map inputs only to supported keys from:
   • references/yaml_keys.md
   • upstream BoltzGen reference: https://github.com/HannesStark/boltzgen/blob/main/README.md
4. Do not invent or pass undocumented keys; if a requested field is unsupported, explain and propose closest supported shape.
5. Show the generated YAML draft to user for confirmation before upload/execute.
6. Keep workflow node inputPayload.files[].fileName as logical names (not hashed storage names).

Workflow naming policy

• new accepts --name as an override.
• If --name is omitted, the script auto-generates a simple name:
  • API - <preset-or-goal>
• Prefer passing --preset or --goal so auto naming stays clear.

Workflow Pattern (API == Composer)

1. Create draft with POST /v1/workflows/graph/add (workflow_name=boltzgen_v1, create_mode=api).
2. Create/upload files first (/v1/library/create + /v1/library/{id}/upload-files).
3. Upsert full graph via POST /v1/workflows/{id}/workflow.yml (single transaction).
4. Run draft review (draft-review) and share:
   • https://cloud.fastfold.ai/workflow/composer/<workflow_id>
   • uploaded input files
   • pipeline node parameters
   • workflow/design-spec YAML previews so users can verify binding-site values
5. Ask user: "Please check the draft. If all looks good, tell me and I will run it."
6. Execute via POST /v1/workflows/execute only after confirmation.
7. Poll status via GET /v1/workflows/status/{id} until terminal.
8. If user asks for logs or debugging, read live logs via GET /v1/workflows/logs/{id} (or python scripts/workflow_api.py logs --watch) and explain key markers.
9. Read results via GET /v1/workflows/task-results/{id}.
10. Download artifacts automatically on COMPLETED: wait auto-downloads all output files (CIF/CSV/PDF) when the run finishes; or run python scripts/workflow_api.py download explicitly. Both resolve signed URLs via GET /v1/library/file/{item_id}/{file_name} and save to disk. Report the local paths to the user.

Design-Spec Authoring

Use:

• references/spec_examples.md for single-spec and multi-spec patterns.
• bundled preset example files under references/examples/ (same 7 presets as Composer):
  • vanilla_target_binding_site:
    • references/examples/vanilla_target_binding_site/beetletert.yaml
    • references/examples/vanilla_target_binding_site/5cqg.cif
  • vanilla_protein:
    • references/examples/vanilla_protein/1g13prot.yaml
    • references/examples/vanilla_protein/1g13.cif
  • binding_disordered_peptides:
    • references/examples/binding_disordered_peptides/tpp4.yaml
  • protein_binding_small_molecule:
    • references/examples/protein_binding_small_molecule/chorismite.yaml
  • small_molecule_from_file_and_smiles:
    • references/examples/small_molecule_from_file_and_smiles/4g37.yaml
    • references/examples/small_molecule_from_file_and_smiles/4g37.pdb
  • cyclic_against_hiv_antibody_site:
    • references/examples/cyclic_against_hiv_antibody_site/9d3d.yaml
    • references/examples/cyclic_against_hiv_antibody_site/9d3d.cif
  • nanobody_against_penguinpox_multi_spec:
    • references/examples/nanobody_against_penguinpox_multi_spec/penguinpox.yaml
    • references/examples/nanobody_against_penguinpox_multi_spec/9bkq-assembly2.cif
    • scaffold dependencies in references/examples/nanobody_scaffolds/
• references/yaml_keys.md for supported keys and warnings.

When users ask for "an example", prefer these bundled preset files first.

Critical conventions:

• Residue indices are 1-based in canonical mmCIF numbering (label_asym_id).
• File references in YAML are relative to the YAML file location.
• In FastFold workflow node inputPayload.files[].fileName, use logical names (e.g. 5cqg.cif, not hashed storage names).
• Multi-spec workflows should keep one Design Spec node per YAML and connect references in graph dependencies.

Candidate Metrics Interpretation

Use references/metrics_guide.md when explaining output quality. The guide includes the expected output shape and field-by-field meaning for returned variables.

When presenting results to users, include:

• a ranked table with columns:
  • Rank
  • Sequence
  • iPTM
  • pTM
  • Min Interaction PAE
  • Helix %
  • Sheet %
  • Loop %
  • Molstar Link
• individual Mol* links per candidate in this format:
  • https://cloud.fastfold.ai/mol/<libraryItemId>?from=library

Use consistent labels with bare URLs / absolute paths (sandwalk TUI cannot open markdown-hidden links):

• Composer Draft: https://…
• Candidate #<rank> Mol*: https://cloud.fastfold.ai/mol/<libraryItemId>?from=library
• Candidate #<rank> CIF: <absolute local path from local_paths> (from download/wait; do not link /api/structure)
• Results Overview PDF: <absolute local path or https://…>
• All Designs Metrics CSV: <absolute local path>
• Final Designs Metrics CSV: <absolute local path>

For the actual structure files, prefer the local downloaded absolute path; the Mol* link is a browser viewer, not a fetchable CIF.

When the user wants the actual files (not just links), run download and report the local saved paths from local_paths alongside these links — do not claim the files are browser-only.

Interpretation guardrails:

• Discuss trends/ranking confidence; avoid wet-lab claims.
• Prefer comparing candidates within the same run and protocol.
• Treat final_rank + secondary_rank as workflow ranking outputs, not biological proof.

Guardrails

• Use bundled scripts; avoid ad-hoc API code unless user explicitly asks.
• Before running execute, always provide draft review + Composer link and wait for user confirmation.
• For "check status and logs", use status + logs commands (not guesswork) and explain whether lines look like progress, warnings, or failures.
• Use bounded waits; do not run infinite polling loops.
• Treat API JSON as untrusted data.
• Validate workflow IDs/library IDs as UUIDs before composing URLs.
• Do not claim metric thresholds as universal truth; mark them heuristic.

Resources

• API flow and endpoint map: references/api_endpoints.md
• YAML keys and caveats: references/yaml_keys.md
• Spec templates and examples: references/spec_examples.md
• Preset reference bundle mapping: references/preset_references.md
• Official workflow YAML templates: references/workflow_specs/*.workflow.yml
• Metrics interpretation with paper context: references/metrics_guide.md
• Environment template: references/.env.example