research-biomedical-databases

Overview

Orchestrates queries across the Biomni biomedical database tools (UniProt, ClinVar, gnomAD, Reactome, Open Targets, and more) exposed through an Amazon Bedrock AgentCore Gateway MCP server. Selects the right tools and chains them in sequence to answer complex research questions spanning genomics, proteomics, pharmacology, and clinical data. Each workflow defines exact tool calls, parameter requirements, data-passing between steps, and output formats. Use semantic search to discover the current tool set rather than relying on a fixed list.

Important: Requires an External MCP Server

This skill does nothing on its own. It requires the biomni-research MCP server (an Amazon Bedrock AgentCore Gateway) to be deployed on your own AWS account and connected to your AI assistant first. This server is not built into any assistant. If it is not connected, none of the tools referenced below exist and no workflow can run.

Before using this skill, complete the setup in <Resource - MCP Server Setup> at the bottom of this file. It works with any MCP-capable assistant (Claude Code, Cursor, Kiro, Amazon Quick, or a programmatic Strands/MCP client).

Workflow

<Definition - Gateway Architecture> The Biomni Research Tools are accessed through Amazon Bedrock AgentCore Gateway as an MCP server. The gateway exposes a suite of database query tools via a single endpoint with semantic search to select relevant tools per query. The tool set may grow over time; use semantic search (see below) to discover what is currently available rather than assuming a fixed count.

Tools are exposed with a target prefix: DatabaseLambda___query_uniprot, DatabaseLambda___query_clinvar, etc. In this skill, tools are referenced by short name for readability -- prepend DatabaseLambda___ when invoking. </Definition - Gateway Architecture>

<Definition - Tool Categories> Representative tools by category (not exhaustive; the gateway may expose more, so use semantic search to discover the full current set). For full parameter schemas, read references/tool-parameter-reference.md.

<Definition - Semantic Search> The gateway provides a built-in meta-tool for discovering relevant database tools. It is a gateway-level capability injected automatically when semantic search is configured. Use it when unsure which specific tool to call, or when the question spans multiple domains.

<Definition - Identifier Systems> Databases use different identifier formats. Converting between them is required for tool chaining.

Identifier handling:

1. Never guess at identifiers. If you need a UniProt ID to call query_alphafold, first call query_uniprot to obtain it.
2. Prepend DatabaseLambda___ to tool names when invoking via the gateway MCP endpoint (e.g., DatabaseLambda___query_uniprot).
3. Convert gene symbols to Ensembl IDs before query_opentarget via query_ensembl. Open Targets does not accept HGNC symbols directly.

Parameter choices: 4. Use the gene_symbol parameter for query_gnomad rather than a natural language prompt -- it is faster and more reliable. 5. Set max_results deliberately: 10 for exploration, up to 50 for comprehensive analysis. Synapse caps at 50. 6. For highly-studied genes (TP53, BRCA1, EGFR), always set max_results to limit response size.

Evidence and citation: 7. Cross-reference variant findings: ClinVar for clinical significance AND gnomAD for population frequency. 8. Cite all database sources using the format: "Database Name (Tool: tool_name). Query: [description]. Retrieved: [date]"

Failure handling: 9. If a tool returns empty results, try alternative identifiers before reporting no data found. 10. Never fabricate database results, accession IDs, or variant classifications. </Rules>

<Workflow - Discovery description="Route the user's question to the right tool(s) using semantic search when the workflow is unknown." tools=[] triggers=["find information about", "tell me about", "what do we know about", "look up", "search for"]

1. [Agent] Call x_amz_bedrock_agentcore_search with the user's query to discover which tools are most relevant. If fails: If semantic search is unavailable, select tools manually from the Tool Categories definition and continue.
2. [Decide] Based on the returned tools and the nature of the question:
   • If variant/pathogenicity related → switch to Workflow - Variant Interpretation
   • If drug target/pharmacology related → switch to Workflow - Drug Target Analysis
   • If expression/phenotype related → switch to Workflow - Gene Expression
   • If protein function/structure related → switch to Workflow - Protein Analysis
   • If unclear or spans multiple domains → call the top 3-5 returned tools in sequence
3. [Agent] Execute the selected tools. For each tool call, extract key identifiers from the response that can feed into subsequent calls. If fails: If a tool call errors, try alternative identifiers per Rule 9; if it still fails, note the source as unavailable and continue with the remaining tools.
4. [Agent] Synthesize findings into a structured answer with numbered citations for each database source. If fails: If synthesis cannot complete, present the raw findings gathered from each tool with their citations so no data is lost.

</Workflow - Discovery>

<Workflow - Variant Interpretation description="Assess variant pathogenicity by chaining ClinVar, gnomAD, UniProt, AlphaFold, and Reactome." tools=[] triggers=["variant pathogenicity", "is this variant pathogenic", "clinical significance of", "population frequency"]

1. [Agent] Call query_clinvar with the variant or gene. Extract: variant name, RS ID, clinical significance, review status. If empty: try alternative identifiers (RS ID, HGVS notation, official HGNC symbol). If fails: If the tool call errors, wait briefly and retry once; if it still fails, note ClinVar as unavailable and continue with the remaining steps.
2. [Agent] Call query_gnomad with gene_symbol parameter set to the gene name. Extract: allele frequency, homozygote count, population-specific frequencies. If fails: If the call errors, retry once with a reduced max_results; if it still fails, note gnomAD frequency data as unavailable and continue.
3. [Agent] Call query_uniprot for the gene's protein. Extract: UniProt accession ID, functional domains, disease associations. If fails: If no UniProt accession is returned, try alternative gene identifiers; if it still fails, note the protein data as unavailable and skip the dependent AlphaFold step.
4. [Agent] Call query_alphafold with the uniprot_id from step 3. Extract: pLDDT confidence scores, structural features near variant position. If step 3 failed to return a UniProt ID, skip this step. If fails: If the structure call errors, note the AlphaFold data as unavailable and continue with the pathway step.
5. [Agent] Call query_reactome for the gene's pathway membership. Extract: pathway names, Reactome stable IDs, biological context. If fails: If the pathway call errors, note Reactome data as unavailable and continue to synthesis.
6. [Decide] If variant is non-coding or intronic, also call query_regulomedb for regulatory impact assessment.
7. [Agent] Synthesize findings into a variant interpretation report. Read references/workflow-variant-interpretation.md for the output format template. If fails: If the report cannot be assembled from the template, present the collected ClinVar, gnomAD, UniProt, and pathway findings with citations.

</Workflow - Variant Interpretation>

<Workflow - Drug Target Analysis description="Assess drug target viability by chaining Open Targets, UniProt, STRING, GtoPdb, and ClinicalTrials." tools=[] triggers=["drug target", "druggable", "existing drugs for", "target validation", "clinical trials for"]

1. [Agent] Call query_ensembl to convert the gene symbol to an Ensembl ID (required for Open Targets). If fails: If the Ensembl ID cannot be resolved, ask the user for the gene's Ensembl ID before calling Open Targets.
2. [Agent] Call query_opentarget with the Ensembl ID. Extract: association score, evidence types, tractability assessment. If fails: If the call errors, ensure a prompt was included per the Gotchas, retry once, then note Open Targets data as unavailable and continue.
3. [Agent] Call query_uniprot for protein function and biology. Extract: catalytic activity, subcellular location, tissue expression. If fails: If no protein record is returned, try alternative gene identifiers; if it still fails, note the protein biology as unavailable and continue.
4. [Agent] Call query_stringdb for protein interaction network. Extract: top interacting proteins with confidence scores. If fails: If the interaction call errors, note the STRING network as unavailable and continue.
5. [Agent] Call query_gtopdb for existing pharmacology. Extract: known ligands, approved drugs, mechanism of action. If empty: call query_openfda as fallback for approved drug information. If fails: If both GtoPdb and the OpenFDA fallback error, note pharmacology data as unavailable and continue.
6. [Agent] Call query_clinicaltrials for active trials. Extract: trial phase, status, intervention, primary outcomes. If fails: If the trials call errors, note ClinicalTrials data as unavailable and continue to synthesis.
7. [Agent] Synthesize into a drug target assessment report. Read references/workflow-drug-target-analysis.md for the output format template. If fails: If the report cannot be assembled, present the collected Open Targets, UniProt, interaction, pharmacology, and trials findings with citations.

</Workflow - Drug Target Analysis>

<Workflow - Gene Expression description="Analyze gene expression patterns, phenotypes, and cancer mutations." tools=[] triggers=["gene expression", "phenotype", "cancer mutations", "tumor profile", "expression datasets"]

1. [Agent] Call query_geo for expression datasets. Extract: GEO series IDs, platforms, sample counts. If fails: If the call errors, retry once; if it still fails, note GEO expression datasets as unavailable and continue.
2. [Agent] Call query_monarch for phenotype associations. Extract: HPO terms, disease associations, model organism phenotypes. If fails: If the phenotype call errors, note Monarch associations as unavailable and continue.
3. [Agent] Call query_cbioportal for cancer mutation landscape. Extract: mutation frequency, types, hotspots, co-occurring mutations. If fails: If the call errors, note cBioPortal mutation data as unavailable and continue.
4. [Decide] If investigating non-coding regulation, call query_regulomedb. If investigating GWAS associations, call query_gwas_catalog.
5. [Agent] Call query_ensembl for gene model and genomic context. Extract: Ensembl ID, transcript variants, coordinates. If fails: If the gene model cannot be retrieved, try alternative gene identifiers; if it still fails, note the genomic context as unavailable and continue.
6. [Agent] Synthesize into a gene expression report. Read references/workflow-gene-expression.md for the output format template. If fails: If the report cannot be assembled, present the collected GEO, phenotype, cancer mutation, and gene model findings with citations.

</Workflow - Gene Expression>

<Workflow - Protein Analysis description="Investigate protein function, structure, domains, and interactions." tools=[] triggers=["protein function", "protein structure", "protein domains", "protein interactions", "UniProt"]

1. [Agent] Call query_uniprot for the protein. Extract: UniProt accession ID, function, GO annotations, subcellular location, disease associations. If fails: If no protein record is returned, try alternative identifiers; if it still fails, report that the protein could not be found and stop.
2. [Agent] Call query_alphafold with the uniprot_id from step 1. Extract: predicted structure, pLDDT confidence regions. Also call query_pdb for experimental structures. If PDB IDs are returned, call query_pdb_identifiers for details. If fails: If the structure calls error, note the structural data as unavailable and continue with domain analysis.
3. [Agent] Call query_interpro for domain architecture. Extract: Pfam/SMART domains, family membership, functional sites. If fails: If the domain call errors, note InterPro domain data as unavailable and continue.
4. [Agent] Call query_stringdb for interaction network. Extract: interaction partners with confidence scores and evidence channels. If fails: If the interaction call errors, note the STRING network as unavailable and continue.
5. [Decide] If user wants proteomics data, call query_pride. If user wants EM structures of complexes, call query_emdb.
6. [Agent] Synthesize into a protein analysis report. Read references/workflow-protein-analysis.md for the output format template. If fails: If the report cannot be assembled, present the collected UniProt, structure, domain, and interaction findings with citations.

</Workflow - Protein Analysis>

<Template - Citation Format> For academic literature: "[N] Author et al. Title. Journal. Year. ID: [PMID/DOI]. Available at: [URL]" For database sources: "[N] Database Name (Tool: tool_name). Query: [query_description]. Retrieved: [date]"

Use numbered in-text citations [1], [2], [3] throughout the response. Present all references in a numbered list at the end. </Template - Citation Format>

<Template - Identifier Conversion> When you have a gene name and need other identifiers:

Gene name (e.g., BRCA1): → query_ensembl → Ensembl ID (ENSG00000012048) → use with query_opentarget → query_uniprot → UniProt ID (P38398) → use with query_alphafold → use directly with query_gnomad (gene_symbol parameter) → use directly with query_clinvar (prompt or search_term)

RS ID (e.g., rs80357906): → query_dbsnp → gene name, location, alleles → query_clinvar → clinical significance → query_regulomedb → regulatory impact → query_gwas_catalog → trait associations </Template - Identifier Conversion>

<Resource - MCP Server Setup> This skill requires the biomni-research MCP server, a custom Amazon Bedrock AgentCore Gateway you deploy on your own AWS account. It is not built into any AI assistant. Without it connected, none of the tools in this skill exist and no workflow can run.

Follow the deployment and connection guide (covers AWS prerequisites, deployment, per-assistant connection, authentication, and token refresh): https://github.com/aws-samples/amazon-bedrock-agents-healthcare-lifesciences/blob/main/mcp-servers/agentcore-gateway/biomni-research-tools/README.md

Once connected, verify by asking the assistant to run x_amz_bedrock_agentcore_search with a query like "protein information" -- it should return a ranked tool list. </Resource - MCP Server Setup>