Biomni Toolkit

A scientific computing skill for working with Biomni — Stanford SNAP Lab's open-source biomedical AI agent framework that autonomously executes complex research tasks across biomedical domains including literature search, data analysis, and experimental design.

When to Use This Skill

Choose Biomni Toolkit when:

• Setting up Biomni agents for automated biomedical research tasks
• Configuring tool collections for literature search, data analysis, or experiment planning
• Building multi-step research workflows with AI-driven hypothesis generation
• Integrating biomedical databases (PubMed, UniProt, ClinicalTrials) through Biomni

Consider alternatives when:

• You need simple PubMed searches (use Biopython Entrez directly)
• You're doing manual data analysis without AI automation
• You need protein structure prediction (use AlphaFold)
• You want a general-purpose AI agent framework (use LangChain or AutoGen)

Quick Start

Copy
claude "Set up a Biomni agent to research drug targets for Alzheimer's disease"

Copy
# Install Biomni
# pip install biomni

from biomni.agent import BioAgent
from biomni.tools import PubMedTool, UniProtTool, DrugBankTool

# Initialize agent with biomedical tools
agent = BioAgent(
    model="gpt-4",
    tools=[
        PubMedTool(),
        UniProtTool(),
        DrugBankTool()
    ],
    verbose=True
)

# Execute a research task
result = agent.run(
    "Find the top 5 novel drug targets for Alzheimer's disease "
    "published in the last 2 years. For each target, provide: "
    "gene name, protein function, supporting evidence, and "
    "any existing drugs in clinical trials."
)

print(result.summary)
for target in result.targets:
    print(f"\n{target.gene}: {target.function}")
    print(f"  Evidence: {target.evidence_count} papers")
    print(f"  Clinical trials: {target.trial_count}")

Core Concepts

Biomni Architecture

Component	Purpose	Example
Agent	Orchestrates research workflow	BioAgent
Tools	Interface to databases and APIs	PubMedTool, UniProtTool
Planner	Decomposes complex queries	Multi-step task planning
Memory	Tracks findings across steps	Accumulated research context
Evaluator	Assesses result quality	Citation verification

Available Tool Collections

Copy
# Literature tools
from biomni.tools import (
    PubMedTool,        # PubMed/MEDLINE search
    BioRxivTool,       # Preprint search
    SemanticScholarTool # Citation analysis
)

# Sequence & structure tools
from biomni.tools import (
    UniProtTool,       # Protein database
    NCBIGeneTool,      # Gene information
    PDBTool,           # Protein structures
    BLASTTool          # Sequence alignment
)

# Clinical tools
from biomni.tools import (
    ClinicalTrialsTool, # Clinical trial registry
    DrugBankTool,       # Drug information
    DisGeNetTool        # Disease-gene associations
)

# Pathway tools
from biomni.tools import (
    KEGGTool,          # Metabolic pathways
    ReactomeTool,      # Biological pathways
    StringDBTool       # Protein interactions
)

Multi-Step Research Workflow

Copy
# Complex research task with multiple phases
workflow = agent.create_workflow([
    {
        "step": "literature_review",
        "query": "Recent CRISPR screens for cancer vulnerabilities",
        "tools": [PubMedTool(), BioRxivTool()],
        "output": "top_papers"
    },
    {
        "step": "target_identification",
        "query": "Extract gene targets from {top_papers}",
        "tools": [UniProtTool(), NCBIGeneTool()],
        "output": "gene_list"
    },
    {
        "step": "druggability_assessment",
        "query": "Assess druggability of {gene_list}",
        "tools": [DrugBankTool(), PDBTool()],
        "output": "druggable_targets"
    }
])

results = workflow.execute()

Configuration

Parameter	Description	Default
model	LLM backend for agent reasoning	gpt-4
max_steps	Maximum agent reasoning steps	10
tools	List of enabled biomedical tools	Required
memory_type	Short-term or long-term memory	short_term
verbose	Print reasoning steps	False

Best Practices

1. Scope research queries precisely. Vague queries like "tell me about cancer" produce broad, unhelpful results. Specify the disease subtype, molecular mechanism, time frame, and desired output format for focused, actionable research summaries.

2. Validate agent findings against primary sources. Biomni agents can hallucinate or misinterpret search results. Always verify critical findings by checking the cited papers directly, especially for claims about clinical trial results or drug efficacy.

3. Use appropriate tool combinations. Different research questions need different tool sets. Literature surveys need PubMed + BioRxiv. Target validation needs UniProt + PDB + DrugBank. Don't load all tools by default — irrelevant tools increase agent confusion.

4. Set reasonable step limits. Complex queries can cause the agent to loop. Start with max_steps=10 and increase only if the agent consistently runs out of steps. If it needs more than 20 steps, the query is probably too broad.

5. Save intermediate results for reproducibility. Export agent reasoning traces and intermediate findings. Biomedical databases update frequently, so the same query may return different results next month. Timestamped exports ensure reproducibility.

Common Issues

Agent returns generic information instead of specific findings. The query is too broad. Add constraints: time period ("published after 2023"), organism ("in human cells"), technique ("using CRISPR"), and output format ("table with gene name, function, evidence level").

Tool calls fail with API rate limits. PubMed and other databases enforce rate limits. Configure retry logic with exponential backoff. For PubMed specifically, register for an NCBI API key to increase the rate limit from 3 to 10 requests per second.

Results include retracted or low-quality papers. Add post-processing filters to exclude retracted papers (check retraction databases), preprints without peer review (if peer-reviewed evidence is needed), and papers from predatory journals. Quality filtering is not automatic.