aws-samples Biomarker Multi-Agent Discovery
When to use this skill
- Complex biomarker discovery requiring multiple analysis modalities
- Coordinating database queries with statistical survival analysis
- Synthesizing findings from literature, pathways, and clinical data into a biomarker panel
- Questions that span multiple sub-domains (e.g., "find best biomarker for survival in chemo patients and show evidence")
Architecture: Agents-as-Tools Pattern
The orchestrator dispatches to specialized sub-agents, each wrapped as a tool:
Orchestrator (Supervisor)
|-- biomarker_database_analyst_agent -> SQL queries on clinical genomic data
|-- clinical_evidence_research_agent -> PubMed + Knowledge Base search
|-- statistician_agent -> Survival regression, Kaplan-Meier plots
|-- medical_imaging_agent -> Radiomics biomarker extraction
Cross-agent data sharing uses AgentCore Memory: the database agent stores query results, and downstream agents (statistician) retrieve them automatically.
Orchestration Workflow
Step 1: Classify the user query
Map the question to required sub-agents:
| Query type | Agents needed | Sequence |
|---|---|---|
| Demographics / counts | Database analyst only | Single call |
| Literature evidence | Clinical evidence researcher only | Single call |
| Statistical analysis (p-values, survival) | Database analyst -> Statistician | Sequential |
| Imaging biomarkers | Database analyst -> Medical imaging | Sequential |
| Comprehensive discovery | All agents | Multi-step |
| Pathway interpretation | Database analyst -> Literature | Sequential |
Step 2: Execute database queries first
For any analysis requiring patient data:
- Call
biomarker_database_analyst_agentwith the data retrieval question - Results are automatically stored in shared memory
- Include required columns: survival_status, survival_duration, biomarker expression values
Example dispatch:
Query: "What are the top 5 biomarkers with overall survival for chemo patients?"
-> Database agent: "Query all records including survival status, survival duration in years, and gene expression values for patients where chemotherapy = 'Yes'"
Step 3: Feed results to downstream agents
For statistical analysis:
-> Statistician agent: "Fit a survival regression model on the query results"
The statistician retrieves data from memory automatically. No S3 path needed.
For visualization:
-> Statistician agent: "Generate a bar chart of the top 5 biomarkers by p-value"
-> Statistician agent: "Plot Kaplan-Meier curve for GDF15 with threshold 10"
For literature validation:
-> Clinical evidence researcher: "Search PubMed for evidence on GDF15 as a biomarker in NSCLC"
Step 4: Synthesize findings into ranked biomarker panel
Combine outputs from all agents into a consolidated report:
Biomarker Panel Report
=====================
1. [Gene] - p-value: X, HR: Y
- Pathway: [from pathway analysis]
- Literature: [N publications supporting]
- Clinical significance: [interpretation]
2. [Gene] - p-value: X, HR: Y
...
Ranking criteria (in priority order):
- Statistical significance (lowest p-value from Cox regression)
- Clinical significance (hazard ratio magnitude)
- Pathway relevance (membership in disease-associated pathways)
- Literature support (number of supporting publications)
- Biological plausibility (protein function matches disease mechanism)
Step 5: Generate actionable recommendations
For each top biomarker, provide:
- Measurement method (IHC, RNA-seq, blood test)
- Patient stratification threshold (expression cutoff)
- Potential clinical utility (prognostic vs. predictive vs. diagnostic)
- Next validation steps (cohort size, assay development)
Tool Dispatch Reference
| Agent | Tools | Input | Output |
|---|---|---|---|
| Database Analyst | get_schema, query_redshift, refine_sql | Data question | Query results (auto-stored in memory) |
| Clinical Evidence | query_pubmed, retrieve | Evidence question | Literature summary with citations |
| Statistician | run_code, plot_kaplan_meier, fit_survival_regression | Analysis request | Regression, charts (S3 paths), p-values |
| Medical Imaging | compute_imaging_biomarker, analyze_imaging_biomarker | Patient IDs | Radiomics features (sphericity, elongation) |
Example Multi-Step Sequences
"Find best biomarker for survival in chemo patients with visualization":
- Database agent -> Statistician (Cox regression) -> Statistician (Kaplan-Meier) -> Literature -> Synthesize
"Compare imaging biomarkers for patients with lowest GDF15":
- Database agent (find patients) -> Medical imaging (compute + visualize) -> Synthesize
Conventions
- Always explain the multi-step plan to the user before executing
- Present results from each agent separately, then provide consolidated summary
- Include S3 paths for any generated charts or images
- When agents fail, explain which step failed and what alternatives exist
- Medical/statistical concepts must be explained in accessible language
- Memory events expire after 3 days -- no manual cleanup needed