Apppublic

joywin2003/crispr-editing-env

sourceHugging Faceupdated 5mo agoView on Hugging Face

0likes

App README

crispr-editing-env v2

A tool-based OpenEnv environment for CRISPR gene editing. An AI agent uses bioinformatics tools to analyze DNA sequences, find PAM sites, design guide RNAs, evaluate off-target risks, and apply precision edits under resource constraints.

Why CRISPR?

CRISPR guide RNA design is a real-world bioinformatics task requiring multi-step investigation: researchers must find PAM sites, design guides, evaluate safety, and manage limited experimental resources. This environment models that workflow as a sequential tool-use problem for LLM agents.

Action Space (8 Bioinformatics Tools)

Tool	Description	Cost
`analyze_sequence <start> <end>`	View GC content, repeats, structure for a gene region	FREE
`search_pam_sites <pattern>`	Find PAM motifs (NGG, NNGRRT) in the gene	FREE
`design_guide <pam_pos> <strand>`	Design a 20nt guide RNA at a PAM site	FREE
`evaluate_guide <sequence>`	Detailed quality scoring	1 credit
`off_target_scan <sequence>`	Check for off-target binding sites	2 credits
`apply_edit <sequence> <position>`	Apply the edit (irreversible)	3 credits
`check_edit_result`	See corrections and damage so far	FREE
`submit_solution`	End episode, trigger grading	FREE

The agent sends free-form tool commands as actions. Budget limits prevent brute-force approaches.

Observation Space

Field	Type	Description
`task_description`	string	What the agent must accomplish
`task_type`	string	singletarget / multirepair / precision_editing
`target_gene_id`	string	Gene identifier
`target_gene_length`	int	Gene length in base pairs
`known_mutations`	list	Mutations to correct (position, refbase, altbase)
`regulatory_regions`	list	No-edit zones (hard task only)
`experiment_budget`	int	Remaining experiment credits
`last_tool_output`	string	Raw output from the last tool call
`tool_history`	list	Previous tool calls and results
`corrections_made`	list	Which mutations have been corrected
`off_target_damage`	list	Any collateral damage from edits

No pre-computed guide candidates or utility scores. The agent must discover everything through tool use.

Tasks (3 Qualitatively Different)

Task	Difficulty	Description	Budget	Max Steps
`single_target`	Easy	Correct one mutation. Find PAM sites, design guide, evaluate, apply.	20	25
`multi_repair`	Medium	Correct 3 mutations with limited budget. Must plan which to group.	15	30
`precision_editing`	Hard	Correct 2 mutations near a regulatory no-edit zone. Obvious guide is a trap.	12	35

Grading

—Easy: 70% correction + 20% safety + 10% efficiency
—Medium: 50% correction + 25% safety + 15% efficiency + 10% grouping bonus
—Hard: 35% correction + 25% safety + 30% regulatory integrity (binary!) + 10% efficiency

Baseline Scores (Greedy Agent, 10 seeds)

single_target       : avg=0.77
multi_repair        : avg=0.63
precision_editing   : avg=0.49

The greedy baseline searches for the nearest PAM site and applies without evaluating off-targets. On the hard task, this frequently damages the regulatory region.

Setup & Usage

Install locally

bash

pip install -r requirements.txt
python inference.py             # LLM inference (needs HF_TOKEN)

Docker

bash

docker build -t crispr-editing-env .
docker run --rm -p 7860:7860 crispr-editing-env

API Endpoints

bash

# Health check
curl http://localhost:7860/

# Reset environment
curl -X POST http://localhost:7860/reset \
  -H "Content-Type: application/json" \
  -d '{"task_level": "single_target", "seed": 42}'

# Take a step
curl -X POST http://localhost:7860/step \
  -H "Content-Type: application/json" \
  -d '{"action": "search_pam_sites NGG"}'

# Get current state
curl http://localhost:7860/state

# List tasks
curl http://localhost:7860/tasks

Environment Variables

bash

export API_BASE_URL="https://router.huggingface.co/v1"
export MODEL_NAME="Qwen/Qwen2.5-72B-Instruct"
export HF_TOKEN="your-token-here"
python inference.py

Project Structure

crispr-editing-env/
├── server/
│   ├── app.py              # FastAPI server (reset/step/state/tasks)
│   ├── environment.py      # CrisprEnv: tool dispatch engine
│   ├── models.py           # Pydantic models
│   ├── simulation.py       # PAM search, guide design, off-target scan
│   ├── tasks.py            # 3 task generators
│   ├── graders.py          # Final scoring logic
│   └── reward.py           # Step-level rewards + tool costs
├── app.py                  # Root re-export for Dockerfile
├── inference.py            # LLM agent with [START]/[STEP]/[END] format
├── openenv.yaml            # OpenEnv spec metadata
├── Dockerfile
├── pyproject.toml
└── requirements.txt