Robin Disease Discovery

Robin — Multi-Agent Disease Discovery System

Robin is FutureHouse's open-source multi-agent system for automating early-stage scientific discovery. Given just a disease name, Robin orchestrates literature search, hypothesis generation, ranking, and (optionally) experimental data analysis to produce a ranked list of:

1. Experimental assays to test pathogenic mechanisms
2. Therapeutic candidates likely to work against the top-ranked assay
3. (Optional) Re-ranked candidates after Finch analyzes any provided wet-lab data

Reported in arXiv:2505.13400 and demonstrated end-to-end on 10 diseases (Glaucoma, Celiac, Friedreich's Ataxia, NASH, etc.) — example outputs ship in the repo.

Prerequisites

• Python ≥ 3.12
• An LLM API key (defaults to o4-mini, set OPENAI_API_KEY; any LiteLLM-compatible provider works)
• An EDISON_API_KEY for the literature/precedent components (Crow / Falcon / Owl) — get one at https://platform.edisonscientific.com/profile
• Wet-lab data analysis step additionally needs Finch access on the platform (skippable)

Install

git clone https://github.com/Future-House/robin.git
cd robin
uv venv .venv && source .venv/bin/activate
uv pip install -e '.[dev]'

cp .env.example .env
# Edit .env to add EDISON_API_KEY and OPENAI_API_KEY (no quotes)

Docker option (recommended for clean environments):

docker build -t robin .
docker run -p 8888:8888 --env-file .env robin
# Open the http://127.0.0.1:8888/lab/tree/robin_demo.ipynb URL Jupyter prints

Run

Either open robin_demo.ipynb in Jupyter, or use programmatically:

from robin import RobinConfiguration, experimental_assay, therapeutic_candidates

config = RobinConfiguration(
    disease_name="Idiopathic Pulmonary Fibrosis",
    # Optional explicit overrides (otherwise read from env):
    # edison_api_key="...",
    # llm_name="claude-opus-4-5",
    num_queries=5,         # literature queries per stage
    num_assays=10,         # candidate assays to generate
    num_candidates=20,     # therapeutic candidates to propose
)

# Stage 1 — generate + rank experimental assays
assay_results = experimental_assay(config)

# Stage 2 — propose + rank therapeutic candidates against the top assay
candidate_results = therapeutic_candidates(config)

# (Optional) Stage 3 — analyze wet-lab data, re-rank
# from robin import data_analysis
# updated = data_analysis(config, dataset_path="my_screen.csv")

What you get back

Robin writes everything to robin_output/<DISEASE>_<TIMESTAMP>/:

robin_output/IPF_2026-05-08_14-30/
├── experimental_assay_summary.txt
├── experimental_assay_detailed_hypotheses/    # one .txt per assay, full reasoning
├── experimental_assay_literature_reviews/      # supporting lit reviews
├── experimental_assay_ranking_results.csv      # pairwise comparison results
├── therapeutic_candidates_summary.txt
├── therapeutic_candidate_detailed_hypotheses/
├── therapeutic_candidate_literature_reviews/
├── ranked_therapeutic_candidates.csv           # final ranked list with scores
└── (if data_analysis run)
    └── data_analysis/                         # Finch outputs, consensus_results.csv

Demo recipe (no wet-lab data)

from robin import RobinConfiguration, experimental_assay, therapeutic_candidates

config = RobinConfiguration(disease_name="Friedreich's Ataxia")
experimental_assay(config)        # ~10–20 min, $5–20 in LLM + Edison credits
therapeutic_candidates(config)    # ~20–40 min
# Open ranked_therapeutic_candidates.csv to see results

The repo also ships examples/<disease>/ — pre-generated outputs for 10 diseases. For a no-credit demo, just open these example folders and walk through the structure.

When to use Robin vs alternatives

Cost / latency expectations

• A full Robin run takes 30 min – 2 hours for typical diseases.
• LLM costs typically $10–$50 per disease with o4-mini defaults; more with larger models.
• Edison credits are consumed by every literature search — budget accordingly.

Caveats

• Robin produces hypotheses to test, not validated science. Treat its ranked candidate list as a starting point for wet-lab work, not as evidence.
• The data_analysis stage requires Edison platform access (Finch). Without it, stages 1–2 still work fully.
• Default LLM is o4-mini — change llm_name in RobinConfiguration for higher-quality runs.
• Examples in examples/ show typical output structure including occasional errors — useful for debugging your own run.