Skip to content

Chaperone

A framework for making structural biology tools interoperable.

Chaperone provides a unified interface for running protein design and prediction tools (RFdiffusion3, AlphaFold3, ProteinMPNN, Boltz-2, and more) as composable blocks with automatic job orchestration, data lineage tracking, and reproducible execution.

  • Simple Example — Minimal pipeline walkthrough (RFD3 → MPNN → AF3)
  • Core Examples — Linked scripts, benchmark objectives, and validation limits
  • Agent Skills — Guides for setup, planning, scripting, and debugging with your coding agent
  • All Blocks — Auto-generated parameter docs for every block, with primary/secondary tier labels
  • Syntax Reference — ExecutionEngine, PipelineDataHandler, JobArgs, RuntimeArgs API reference
  • Validation — Score any structure predictor on the PDB validation cohort
  • Evaluations — Run the agent evaluation harness on the design tasks

Getting Started

# Install
uv venv --python=3.12.3
source .venv/bin/activate
uv pip install -e '.[all]'

# Check which blocks and containers are available
chaperone list-blocks

Architecture

Each pipeline follows a composable block pattern:

  1. Backbone generation (rfd3, rfd1, disco) — generate diverse backbones
  2. Filtering (filter) — prune low-quality backbones early
  3. Sequence design (mpnn, foundry_mpnn, abmpnn_poe) — inverse fold sequences
  4. Structure prediction (af3, rf3, boltz2) — validate with a prediction oracle
  5. Metrics + filtering (rmsd, rasa, hbond, filter) — assess quality and select top designs

Advanced patterns include cycling (iterative MPNN → oracle loops), parallel submission (engine.submit() + gather()), and branching (multiple prediction configurations).

See the Simple Example for a pipeline walkthrough and Agent Skills for guided workflows. The /setup guide covers containers and compute. Skills are Markdown instructions that coding agents can read; automatic discovery and slash commands depend on the agent installation. They still require working containers, weights, and compute access.

Chaperone-authored code is distributed under the MIT license. Bundled third-party code and example outputs, upstream tools, and model weights retain their own terms and access requirements. See the installation instructions to configure your own system, or supply an untracked .env.ipd for an IPD installation.