FLASH-P

A multi-agent system that turns primary literature into validated, perturbation-testable causal networks — for any trait, in any species.

TL;DR  ·  What it is  ·  Versions  ·  How to run  ·  Commands  ·  Repo  ·  GUI

---

TL;DR — run it in 4 steps

New here? This is the whole thing:

1. Download the repo — green Code ▸ Download ZIP at the top of this page (or git clone), then unzip it. Pick your version folder — start with Flash-P_Plant/.
2. Open a terminal in the Claude subfolder of that version — i.e. inside Flash-P_Plant/Claude/ — and start Claude Code there (run claude). That subfolder is the project root Claude Code should open.
3. Run a command — e.g. /run-flashp <trait> in <species>. Don't want to approve every step? Press shift + tab to turn on auto-accept and let it run.
4. Go do something fun ☕ — come back to a finished, DOI-backed network. Then /run-flashp-studio networks to browse, perturb, and export it.

Want the details? Keep reading. New to the commands? Run /run-flashp-help inside Claude Code.

---

What is FLASH-P?

FLASH-P reads primary scientific literature straight from the web and turns it into a validated, perturbation-testable causal network for any (trait, species) pair. A sequence of specialised agents — literature reviewer → judge → builder → judge → perturbation → validator → refiner → exporter — pass strict handoff files to each other. Every edge carries a DOI; every node is wired into algebraic, Hill-function ODE, and signed Random-Walk-with-Restart equations, so the network is both human-readable and numerically simulable.

See it for yourself at https://flash-p.com/ — browse the networks, run perturbations, and explore the outcomes. Click through to get a feel for what FLASH-P builds.

The agents are just plain-text prompts plus structured handoff files — nothing in them is tied to one model or tool. The same pipeline runs inside any coding agent that reads a CLAUDE.md or AGENTS.md file — we ship ready-to-go folders for Claude Code, Codex CLI, and OpenCode (which also covers Aider, Goose, and any other AGENTS.md-aware tool).

FLASH-P has already built hundreds of networks. The 13 shipped in Outcome/ are the ones from the paper — built with the paper version of the agents: six Arabidopsis phenotype networks, one merged six-trait network, and six other-species networks (E. coli, maize, poplar, rice, sorghum, wheat). On a like-for-like benchmark they reach higher direction-call accuracy than a cleaned knowledge-base baseline at a fraction of the size — see Outcome/FLASH-P_VS_KG/.

---

Versions

Pick the folder for your domain, then open the subfolder for your coding agent.

Version	Use it for	Notes
Flash-P_Plant/	Plant & crop traits	Start here. The lean, current pipeline, tested most extensively. Ships Claude Code, Codex, and OpenCode variants.
Flash-P_Medical/	Medical traits	🚧 Alpha. Same pipeline, medical tuning. Works, but not yet tested as extensively as Plant. Claude Code setup.
Flash-P_Animal/	Animal traits	🚧 Alpha. Same pipeline, animal tuning. Works, but not yet tested as extensively as Plant. Claude Code setup.
Flash-P_Paper_Version/	Reproducing the manuscript	The original pipeline behind the paper — full provenance and multi-pass judges, so heavy on tokens. Kept for reproducibility. The current versions above are far leaner with the same science.

In short: the current versions are very lean and fit a normal session; the paper version is the heavier original, kept only so the published results can be reproduced.

---

How to run

Open the version folder for your domain, then the subfolder for your coding agent, and run the pipeline for a trait and species (e.g. Shoot Branching in Arabidopsis).

Claude Code

# install agentic coding platform (e.g. Claude Code)
npm install -g @anthropic-ai/claude-code   # or get the desktop app: https://www.claude.com/product/claude-code

# Download Flash-P from github. No actual installation required
git clone https://github.com/CMits/FlashP.git

# run your agent platform from the correct directory within Flash-P
cd FlashP/Flash-P_Plant/Claude
claude

Then run:

/run-flashp Shoot Branching in Arabidopsis

It runs the whole pipeline (Steps 1→6) autonomously and writes the network, validation, and supplementary tables into a new networks/<trait>/ folder.

Codex / OpenCode / Aider / Goose

Open the matching subfolder — Flash-P_Plant/Codex/ or Flash-P_Plant/OpenCode_Aider_Any_Other/ — start your agent there, and paste:

Run the full FLASH-P pipeline for <trait> in <species>.

Open the agent subfolder, not the repo root — the pipeline reads its orchestrator file (CLAUDE.md / AGENTS.md) and Agent/*.md as relative paths.

Medical & Animal currently ship the Claude Code setup (CLAUDE.md + Agent/); open that folder in Claude Code and use the same /run-flashp command.

---

Commands

Once your agent is running in a version folder (e.g. Flash-P_Plant/Claude), these slash commands are available. Run /run-flashp-help at any time to print this list from inside Claude Code.

Command	What it does	Usage
/run-flashp	Autonomously run the full pipeline (Steps 1→6) for a trait and write the network, validation, and supplementary tables into networks/<trait>/.	/run-flashp <trait> in <species>
/run-flashp-studio	Build the FLASH-P Studio — one self-contained, offline HTML app to browse, view (clickable, DOI links), export each network as PNG/SVG, and perturbate all your networks (KO/KD/OE + treatments; Algebraic / RWR / ODE solvers, live charts).	/run-flashp-studio <networks dir>
/run-flashp-epistasis	Gene × gene epistasis scan over an existing network — every single + double perturbation, classified by interaction. (Plant)	/run-flashp-epistasis <network dir>
/run-flashp-gxe	Gene × environment (G×E) scan over an existing network — dose-swept, with a report. (Plant)	/run-flashp-gxe <network dir>
/run-flashp-help	Print this command list from inside Claude Code (auto-generated from the commands available in the folder).	/run-flashp-help

/run-flashp, /run-flashp-studio, and /run-flashp-help ship in Plant, Animal, and Medical; /run-flashp-epistasis and /run-flashp-gxe currently ship in the Plant variant.

---

What's in this repo

Folder	Contents
Flash-P_Plant/, Flash-P_Medical/, Flash-P_Animal/	The latest versions (see Versions).
Flash-P_Paper_Version/	The original heavy-token pipeline used for the manuscript.
Outcome/	All paper outputs: every network, validation result, refinement history, Cytoscape export, the KG-baseline comparison, and a local open-source-model reproduction.
Supplementary Data/	Seven supplementary .xlsx tables with build scripts and per-dataset descriptions.
Images/	Banner, pipeline diagram, and GUI preview.

Visualise in Cytoscape

Every network in Outcome/Networks/ ships as .graphml — drop one onto Cytoscape and import Style_Cytoscape.xml (style FLASH-P) to render it with the published palette. Full recipe in Outcome/Networks/README.md.

---

FLASH-P GUI — coming soon

A no-install desktop app is in development: build, browse, edit, save, and merge networks; run AI-driven analyses across many networks at once; and chat with your networks in natural language — all with a built-in model picker (OpenAI, Anthropic, Gemini, Kimi, DeepSeek, Qwen, or any local model). Stay tuned at https://flash-p.com/. Status: not yet released.

---

Creators

Christos Mitsanis · David Kainer — The University of Queensland

Contact

• Christos Mitsanis — c.mitsanis@uq.edu.au
• David Kainer — d.kainer@uq.edu.au
• Or open a GitHub issue.

License

© 2026 The University of Queensland, released under CC BY-NC-SA 4.0 — free to share and adapt for non-commercial use with attribution and share-alike. If you use FLASH-P in academic work, please cite:

Mitsanis C, Fortuna NZ, Beveridge C, Kainer D. Turning decades of biology into accurate causal networks with AI agents. bioRxiv 2026.06.13.731799; doi: 10.64898/2026.06.13.731799