BioSymphony Small Molecules

BioSymphony Small Molecules is an agent skill for choosing open tools in small-molecule design.

Agent skill for routing small-molecule design tasks to open tools, focused references, and license-aware implementation paths.

It gives Claude Code, Codex, Symphony, and other agent harnesses a structured map for:

• synthesizable molecule generation and analog design
• retrosynthesis, reaction prediction, templates, and makeability scoring
• protein structure prediction, docking, co-folding, and binding affinity
• QSAR, ADMET, selectivity, and pocket-conditioned generation
• code, weights, data, and base-model license terms

The goal is practical routing. An agent can read the skill, pick the right reference file, choose a starting tool, and keep the licensing and data constraints visible while it works through a long design task.

What This Improves

Agent need	What this repo adds
Pick a starting method	Task routing in SKILL.md and the compact tool matrix
Avoid loading too much context	Focused reference files by workflow category
Compare tools quickly	Tool cards with task fit, license notes, weights, data, and status
Connect target work to synthesis	A two-layer loop from target scoring back to makeable molecules
Keep public demos small	A compact KRAS molecular-glue example with public inputs and summaries

Agent Flow

flowchart TD
    A["User task"] --> B["Open SKILL.md"]
    B --> C["Identify task category"]
    C --> D["Load one focused reference"]
    D --> E["Choose first tool path"]
    E --> F["Check code, weights, data, and base-model terms"]
    F --> G["Run or recommend the smallest suitable workflow"]
    G --> H["Record limits and next decision"]

Loading

Design Loop

flowchart LR
    A["Target question"] --> B["Structure, docking, co-folding, affinity"]
    B --> C["QSAR, ADMET, selectivity"]
    C --> D["Candidate set"]
    D --> E["Synthesizable projection and analogs"]
    E --> F["Retrosynthesis and route checks"]
    F --> G["Makeable candidates"]
    G --> B
    F --> H["License and data review"]
    B --> H

Loading

How Agents Use It

Open SKILL.md. The skill routes by task:

• generate makeable analogs of a hit
• project a molecule into synthesizable space
• plan or check a synthesis route
• dock a ligand or co-fold a protein-ligand complex
• estimate binding affinity with an ML or free-energy method
• build QSAR, ADMET, or off-target screens
• generate molecules into a binding pocket
• check whether a tool can be used in commercial work

The reference files are plain Markdown. They work as agent context and as human-readable notes.

For Claude Code-style skill discovery:

ln -s "$(pwd)" ~/.claude/skills/small-molecule-design-tools

The skill name is small-molecule-design-tools.

What Is Included

SKILL.md                         agent entry point and routing table
references/tool-matrix.md        compact index of tools by task
references/licensing-and-data.md code, weights, data, and base-model checklist
references/*.md                  focused tool cards by category
demos/kras-glue/                 compact public-data demo on PDB 9BG6
assets/readme-banner.png         curated README banner image
scripts/public_audit.py          public-release scan for local paths, secrets, and links

The reference layer covers about 130 tools. The synthesizability layer was checked on 2026-06-11. The target-based layer was checked on 2026-06-13.

Start Points

Start here	Use it for
references/tool-matrix.md	One table across all tool categories
references/licensing-and-data.md	Code, weights, data, and base-model terms
references/synthesizable-generation.md	Makeable molecule generation and analog design
references/retrosynthesis-planning.md	Multi-step synthesis planning
references/docking-and-cofolding.md	Docking, co-folding, pose, and affinity tools
references/binding-affinity-and-fep.md	OpenFE, OpenMM, MM-GBSA, and related methods
references/worked-example-kras-glue.md	Applying the layers to a public KRAS molecular-glue structure

Public Repo Boundary

This repo contains docs, skill instructions, compact public-data examples, and small result summaries. It does not include model weights, vendor catalogs, non-public molecules, non-public structures, raw cloud outputs, or large media builds.

Run the public-release check:

make release-check

This compiles the public Python scripts, checks local Markdown links, and scans for local workstation paths, secrets, and oversized files.

License

The repo content is MIT licensed. Upstream tools keep their own licenses. The reference cards separate source code, model weights, data, and base-model terms because those terms often differ.