biomni（biomni 搜索）

Biomni

Overview

Biomni is an open-source biomedical AI agent framework from Stanford's SNAP lab that autonomously executes complex research tasks across biomedical domains. Use this skill when working on multi-step biological reasoning tasks, analyzing biomedical data, or conducting research spanning genomics, drug discovery, molecular biology, and clinical analysis.

Core Capabilities

Biomni excels at:

1. Multi-step biological reasoning - Autonomous task decomposition and planning for complex biomedical queries
2. Code generation and execution - Dynamic analysis pipeline creation for data processing
3. Knowledge retrieval - Access to ~11GB of integrated biomedical databases and literature
4. Cross-domain problem solving - Unified interface for genomics, proteomics, drug discovery, and clinical tasks

When to Use This Skill

Use biomni for:

• CRISPR screening - Design screens, prioritize genes, analyze knockout effects
• Single-cell RNA-seq - Cell type annotation, differential expression, trajectory analysis
• Drug discovery - ADMET prediction, target identification, compound optimization
• GWAS analysis - Variant interpretation, causal gene identification, pathway enrichment
• Clinical genomics - Rare disease diagnosis, variant pathogenicity, phenotype-genotype mapping
• Lab protocols - Protocol optimization, literature synthesis, experimental design

Quick Start

Installation and Setup

Install Biomni and configure API keys for LLM providers:

uv pip install biomni --upgrade

Configure API keys (store in .env file or environment variables):

export ANTHROPIC_API_KEY="your-key-here"
# Optional: OpenAI, Azure, Google, Groq, AWS Bedrock keys

Use scripts/setup_environment.py for interactive setup assistance.

Basic Usage Pattern

from biomni.agent import A1

# Initialize agent with data path and LLM choice
agent = A1(path='./data', llm='claude-sonnet-4-20250514')

# Execute biomedical task autonomously
agent.go("Your biomedical research question or task")

# Save conversation history and results
agent.save_conversation_history("report.pdf")

Working with Biomni

1. Agent Initialization

The A1 class is the primary interface for biomni:

from biomni.agent import A1
from biomni.config import default_config

# Basic initialization
agent = A1(
    path='./data',  # Path to data lake (~11GB downloaded on first use)
    llm='claude-sonnet-4-20250514'  # LLM model selection
)

# Advanced configuration
default_config.llm = "gpt-4"
default_config.timeout_seconds = 1200
default_config.max_iterations = 50

Supported LLM Providers:

• Anthropic Claude (recommended): claude-sonnet-4-20250514, claude-opus-4-20250514
• OpenAI: gpt-4, gpt-4-turbo
• Azure OpenAI: via Azure configuration
• Google Gemini: gemini-2.0-flash-exp
• Groq: llama-3.3-70b-versatile
• AWS Bedrock: Various models via Bedrock API

See references/llm_providers.md for detailed LLM configuration instructions.

2. Task Execution Workflow

Biomni follows an autonomous agent workflow:

# Step 1: Initialize agent
agent = A1(path='./data', llm='claude-sonnet-4-20250514')

# Step 2: Execute task with natural language query
result = agent.go("""
Design a CRISPR screen to identify genes regulating autophagy in
HEK293 cells. Prioritize genes based on essentiality and pathway
relevance.
""")

# Step 3: Review generated code and analysis
# Agent autonomously:
# - Decomposes task into sub-steps
# - Retrieves relevant biological knowledge
# - Generates and executes analysis code
# - Interprets results and provides insights

# Step 4: Save results
agent.save_conversation_history("autophagy_screen_report.pdf")

3. Common Task Patterns

CRISPR Screening Design

agent.go("""
Design a genome-wide CRISPR knockout screen for identifying genes
affecting [phenotype] in [cell type]. Include:
1. sgRNA library design
2. Gene prioritization criteria
3. Expected hit genes based on pathway analysis
""")

Single-Cell RNA-seq Analysis

agent.go("""
Analyze this single-cell RNA-seq dataset:
- Perform quality control and filtering
- Identify cell populations via clustering
- Annotate cell types using marker genes
- Conduct differential expression between conditions
File path: [path/to/data.h5ad]
""")

Drug ADMET Prediction

agent.go("""
Predict ADMET properties for these drug candidates:
[SMILES strings or compound IDs]
Focus on:
- Absorption (Caco-2 permeability, HIA)
- Distribution (plasma protein binding, BBB penetration)
- Metabolism (CYP450 interaction)
- Excretion (clearance)
- Toxicity (hERG liability, hepatotoxicity)
""")

GWAS Variant Interpretation

agent.go("""
Interpret GWAS results for [trait/disease]:
- Identify genome-wide significant variants
- Map variants to causal genes
- Perform pathway enrichment analysis
- Predict functional consequences
Summary statistics file: [path/to/gwas_summary.txt]
""")

See references/use_cases.md for comprehensive task examples across all biomedical domains.

4. Data Integration

Biomni integrates ~11GB of biomedical knowledge sources:

• Gene databases - Ensembl, NCBI Gene, UniProt
• Protein structures - PDB, AlphaFold
• Clinical datasets - ClinVar, OMIM, HPO
• Literature indices - PubMed abstracts, biomedical ontologies
• Pathway databases - KEGG, Reactome, GO

Data is automatically downloaded to the specified path on first use.

5. MCP Server Integration

Extend biomni with external tools via Model Context Protocol:

# MCP servers can provide:
# - FDA drug databases
# - Web search for literature
# - Custom biomedical APIs
# - Laboratory equipment interfaces

# Configure MCP servers in .biomni/mcp_config.json

6. Evaluation Framework

Benchmark agent performance on biomedical tasks:

from biomni.eval import BiomniEval1

evaluator = BiomniEval1()

# Evaluate on specific task types
score = evaluator.evaluate(
    task_type='crispr_design',
    instance_id='test_001',
    answer=agent_output
)

# Access evaluation dataset
dataset = evaluator.load_dataset()

Best Practices

Task Formulation

• Be specific - Include biological context, organism, cell type, conditions
• Specify outputs - Clearly state desired analysis outputs and formats
• Provide data paths - Include file paths for datasets to analyze
• Set constraints - Mention time/computational limits if relevant

Security Considerations

⚠️ Important: Biomni executes LLM-generated code with full system privileges. For production use:

• Run in isolated environments (Docker, VMs)
• Avoid exposing sensitive credentials
• Review generated code before execution in sensitive contexts
• Use sandboxed execution environments when possible

Performance Optimization

• Choose appropriate LLMs - Claude Sonnet 4 recommended for balance of speed/quality
• Set reasonable timeouts - Adjust default_config.timeout_seconds for complex tasks
• Monitor iterations - Track max_iterations to prevent runaway loops
• Cache data - Reuse downloaded data lake across sessions

Result Documentation

# Always save conversation history for reproducibility
agent.save_conversation_history("results/project_name_YYYYMMDD.pdf")

# Include in reports:
# - Original task description
# - Generated analysis code
# - Results and interpretations
# - Data sources used

Resources

References

Detailed documentation available in the references/ directory:

• api_reference.md - Complete API documentation for A1 class, configuration, and evaluation
• llm_providers.md - LLM provider setup (Anthropic, OpenAI, Azure, Google, Groq, AWS)
• use_cases.md - Comprehensive task examples for all biomedical domains

Scripts

Helper scripts in the scripts/ directory:

• setup_environment.py - Interactive environment and API key configuration
• generate_report.py - Enhanced PDF report generation with custom formatting

External Resources

• GitHub: https://github.com/snap-stanford/biomni
• Web Platform: https://biomni.stanford.edu
• Paper: https://www.biorxiv.org/content/10.1101/2025.05.30.656746v1
• Model: https://huggingface.co/biomni/Biomni-R0-32B-Preview
• Evaluation Dataset: https://huggingface.co/datasets/biomni/Eval1

Troubleshooting

Common Issues

Data download fails

# Manually trigger data lake download
agent = A1(path='./data', llm='your-llm')
# First .go() call will download data

API key errors

# Verify environment variables
echo $ANTHROPIC_API_KEY
# Or check .env file in working directory

Timeout on complex tasks

from biomni.config import default_config
default_config.timeout_seconds = 3600  # 1 hour

Memory issues with large datasets

• Use streaming for large files
• Process data in chunks
• Increase system memory allocation

Getting Help

For issues or questions:

• GitHub Issues: https://github.com/snap-stanford/biomni/issues
• Documentation: Check references/ files for detailed guidance
• Community: Stanford SNAP lab and biomni contributors