BioSymphony Structure Factory helps scientists and AI agents select, call, and chain structural biology tools. It combines a tool knowledge base with a CLI, agent skills, and an adapter executor for design, structure prediction, and model comparison.
Use tool cards to choose a method, bind an installed tool or service client, and pass its checked outputs to the next stage. Scientific tools, model weights, and provider access are configured separately.
| Capability | What You And Your Agent Can Do | Start With |
|---|---|---|
| Find tools | Compare roles, inputs, outputs, runtime needs, licenses, source repos, and papers | Tool cards, software registry, tool radar |
| Call tools | Invoke a configured local program or an API, cloud, scheduler, or container client through an adapter | Adapter guide |
| Chain calls | Connect stage dependencies and pass checked files or directory bundles into downstream tools | Controller guide |
| Compare methods | Hold inputs and scoring fixed across toolchain arms; retain failed candidates in the comparison | Binder comparison guide, scoring stack |
| Add a tool | Supply an adapter with typed inputs and expected outputs, or use an installed platform skill | Custom tools and routes |
| To | Start With |
|---|---|
| Give your agent a scientific task | Structure Factory skill and example request |
| Choose a tool or model | Tool cards and tool radar |
| Call tools and connect stages | Execution guide |
| Inspect the CLI locally | Install and inspect |
Give your agent the Structure Factory skill and describe the result you need.
Use the BioSymphony Structure Factory skill.
Goal: <design candidates, predict structures, compare models, screen, or render>.
Inputs: <public accessions or synthetic fixtures>.
Tools: <choose from the knowledge base, or use these named tools>.
Workflow: <which tool outputs feed which later steps>.
Compute: <local tools, installed skills, or selected hosted services>.
Limits: <budget, runtime, candidate count, metric, and stopping rule>.
Results: <structures, score tables, candidate rankings, or figures>.
Check tool availability and input compatibility, then dry-run the toolchain.
Identify any missing adapter, installation, or authorization before execution.
The public CLI uses bsf binder-lane adapter for one tool call and
bsf binder-lane execute for a prepared adapter chain. Its run command
exercises the included synthetic fixture. Installed platform skills are called
by your agent, then checked against the declared stage outputs.
Obtain explicit human authorization before a paid provider start, non-public upload, terms acceptance, or large or license-gated download. The approval names the route, data posture, budget, runtime, and applicable terms.
Use the Structure Factory skill to compare structure-prediction tools for public inputs. Read the tool cards, explain which methods fit the inputs, and configure separate prediction arms with the same review criteria. Connect prediction outputs to scoring and visualization. Return a comparison table, failed-input list, and selected figures. Dry-run the calls before execution.
Use cases includes requests for binder comparison, state atlases, screening, and structure mapping.
Select tools and execution routes for the stages you need. The binder controller runs configured adapters in dependency order and verifies file hashes and output counts before passing results downstream. A failed tool or invalid handoff stops the remaining stages and returns a failure record with next actions.
A tool card supplies method knowledge. A runnable adapter supplies the command and input/output mapping for your installation. Check both when building a chain; the execution guide explains how to resolve missing bindings and connect installed skills.
Use Structure Factory when a scientific question needs several tools, a comparison between methods, or an agent that can inspect and continue the work. Choose the steps that answer the question: prediction and review, design and scoring, screening, or structural visualization.
Inspect command bindings with
bsf binder-lane adapters, then dry-run the selected adapter to check your
installation.
| Role | Example Tool Records | Output To Use Downstream |
|---|---|---|
| Generate candidate structures | RFdiffusion3, Genie3, BindCraft2 | Candidate structures or designs |
| Design sequences | ProteinMPNN | Sequences for selected backbones |
| Predict complexes | Boltz, Chai, ESMFold2 | Predicted structures and confidence fields |
| Score and filter | Cofold scoring stack, refinement stack | Per-tool scores, filter decisions, and failure rows |
| Review poses and screens | PoseBusters, MolPAL | Pose checks and selected screening candidates |
| Render structures | ChimeraX, MolViewSpec | Structural figures and reusable molecular views |
The bundled execution registry includes command records for Boltz, ESMFold2 full/Fast, supplied-backbone input, and status and diversity filters. Other tool records may need an installation-specific adapter or platform skill. See capabilities for execution coverage and the tool radar for dated research leads.
Use the Markdown skills with Codex, Claude Code, or another agent that reads
files and calls commands. Use bsf directly when you prefer a terminal.
A toolchain can mix local programs, hosted APIs, installed platform skills, and compute on RunPod, FAL, Modal, Lambda Cloud, AWS, or SSH/HPC. Configure a matching adapter or skill for each stage. Provider profiles record setup requirements; the selected route determines runtime availability.
Optional task packs let OpenAI's Symphony with Linear, GitHub Issues, or another queue coordinate the same work.
Install the CLI and inspect tools and adapters. These commands run locally without a GPU or provider account:
git clone https://github.com/BioSymphony/structure-factory.git
cd structure-factory
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e .
bsf doctor .
bsf binder-lane menu --workspace .
bsf binder-lane adapters --workspace .
bsf catalog . --format markdown
bsf validate examples/pd-l1-binder-design-publicFollow the execution guide to bind inputs, dry-run one adapter, or prepare and execute a chain. Planning and validation need Python 3.10+; each scientific tool has its own runtime requirements.
Campaign scaffolding and optional tracker tasks
bsf scaffold-campaign .runtime/pd-l1-binder-demo \
--campaign-id pd-l1-binder-demo \
--target-label "PD-L1 public interface demo" \
--public-accession "PDB:4ZQK" \
--window "public PD-1/PD-L1 interface window"
bsf validate .runtime/pd-l1-binder-demo
bsf issue-dry-run examples/pd-l1-binder-design-public \
--out .runtime/pd-l1-issuesThe workflow map and agent harness guide cover task dependencies and worker handoffs.
The PD-L1 example includes a target
window for PDB 4ZQK, generation and cofold plans, and a candidate ranking
schema. The fast-path recipe covers
local preparation. Verify coordinate inputs with target-check --plan <plan.json>
before a comparative generation run.
The binder-lane skill also supports an independent comparison workflow based on Anthropic's public report and released dataset. Choose source-tool replay, workflow-shape replay, or deliberate tool swaps.
- FAQ: GPUs, agents, trackers, and adding tools.
- Glossary: scientific and workflow terms.
- Quickstart tour: a local CLI walkthrough.
- CLI reference: commands and options.
- Skill installation: agent setup.
Operations, repository layout, and release checks
Read the preflight checklist before paid execution. Operational checks and output validation cover failed tools, incomplete results, and cleanup.
| Directory | Contents |
|---|---|
tools/, references/ |
Tool knowledge base, registries, and source records |
skills/, src/ |
Agent instructions and the bsf CLI |
recipes/, examples/, demos/ |
Workflows, fixtures, and curated examples |
campaigns/, modules/, schemas/ |
Campaign definitions and reusable stage contracts |
runpod/, packs/, templates/ |
Provider templates and optional task coordination |
scripts/, tests/, docs/ |
Validators, tests, and guides |
Before publishing, follow PUBLIC_RELEASE.md and run
make public-switch-check, then make clean. Keep credentials, private data,
provider logs, model weights, and generated structures in runtime storage.
Pre-alpha. Local checks cover planning, adapter execution, stage handoffs, fixtures, and release scans. Scientific tool availability depends on the configured installation or service. NON_CLAIMS.md and BIOSAFETY.md define the result boundaries.