openai/locus-to-gene-mapper-skill

Locus-to-Gene Mapper

Generate a reproducible locus-to-gene mapping for one trait (or a seed set of lead variants), with explicit evidence attribution and conservative confidence labels.

This skill is optimized for bioinformaticians who need executable, traceable mapping from variant signals to plausible causal genes.

Required Inputs

Provide at least one anchor source:

• trait_query (string), for example chronic obstructive pulmonary disease
• efo_id (string), for example EFO_0000341
• seed_rsids (list[string]), for example ["rs1873625", "rs7903146"]

Optional Inputs

• target_gene (string), optional gene of interest for highlighting in output
• show_child_traits (bool), default true
• phenotype_terms (list[string]), optional additional terms to include when finding anchors
• max_anchor_associations (int), default 1200
• max_loci (int), default 25
• max_genes_per_locus (int), default 10
• max_coloc_rows_per_locus (int), default 100
• max_eqtl_rows_per_variant (int), default 200
• genebass_burden_sets (list[string]), default ["pLoF", "missense|LC"]
• include_clinvar (bool), default true
• include_gnomad_context (bool), default true
• include_hpa_tissue_context (bool), default true
• include_figures (bool), default false
• disable_default_seeds (bool), default false; if false, common traits automatically get built-in seed rsIDs
• figure_output_dir (string), default ./output/figures
• mapping_output_path (string), default ./output/locus_to_gene_mapping.json
• summary_output_path (string), default ./output/locus_to_gene_summary.md

Runtime Requirements

• Python 3.11+
• requests
• Optional for figure generation: matplotlib, seaborn, pandas

Bundled Script (Deterministic Runner)

• Primary entrypoint: scripts/map_locus_to_gene.py
• This script:
  • resolves trait/EFO and anchor variants,
  • resolves seed and anchor rsID coordinates directly through NCBI RefSNP/dbSNP placements,
  • gathers locus-to-gene evidence through the chained skills,
  • writes mapping JSON and summary markdown,
  • optionally renders figures when plotting deps are available.

Run:

python locus-to-gene-mapper-skill/scripts/map_locus_to_gene.py \
  --input-json /path/to/input.json \
  --print-result

Quick start (no input JSON file):

python locus-to-gene-mapper-skill/scripts/map_locus_to_gene.py \
  --trait-query "type 2 diabetes" \
  --print-result

Trait-only runs default to include_figures=true unless explicitly disabled with --no-include-figures.

Minimal input JSON:

{
  "trait_query": "type 2 diabetes"
}

Built-in default seeds (when disable_default_seeds=false):

• type 2 diabetes / t2d -> rs7903146, rs13266634, rs7756992, rs5219, rs1801282, rs4402960
• coronary artery disease / cad -> rs1333049, rs4977574, rs9349379, rs6725887, rs1746048, rs3184504
• body mass index / bmi -> rs9939609, rs17782313, rs6548238, rs10938397, rs7498665, rs7138803
• asthma -> rs7216389, rs2305480, rs9273349
• rheumatoid arthritis -> rs2476601, rs3761847, rs660895
• alzheimer disease -> rs429358, rs7412, rs6733839, rs11136000, rs3851179
• ldl cholesterol / total cholesterol -> rs7412, rs429358, rs6511720, rs629301, rs12740374, rs11591147

Autonomous Execution Contract (Embedded Behavior)

When a user asks for locus-to-gene mapping and gives only a trait (for example, type 2 diabetes), do the following automatically:

1. Run the bundled script with --trait-query "<user_trait>" --print-result (no manual JSON required).
2. If it returns No anchors remained, rerun once with a built-in default seed rsID for that trait (unless disable_default_seeds=true).
3. Read the generated mapping_output_path and summary_output_path.
4. Return this concise response structure:
   • Top 5 cross-locus prioritized genes
   • Per-locus top gene (score, confidence)
   • Visualization artifact (figure path(s) or Mermaid fallback block)
   • Warnings and limitations
5. For inline image rendering in chat:
   • read inline_image_markdown from script result
   • emit those lines exactly as plain markdown (no code fences)
   • if inline rendering still fails, instruct user to upload PNG files into the chat

Do not ask the user to run python manually unless execution is actually blocked.

Skill Chaining Order (Mandatory)

Use these skills in order. Skip only when an earlier step is not needed by provided inputs.

1. efo-ontology-skill
   • Resolve trait_query to canonical EFO term and synonyms.
   • Expand descendants when show_child_traits=true.
2. gwas-catalog-skill
   • Discover anchor variants for the trait/EFO scope.
   • Pull association/study metadata for locus context.
3. Built-in NCBI RefSNP coordinate resolution
   • Normalize each anchor rsID to GRCh37/GRCh38 top-level chromosome placements.
4. opentargets-skill
   • Retrieve credible set context, L2G predictions, and colocalisation evidence per locus.
5. gtex-eqtl-skill
   • Retrieve single-tissue eQTL support for anchor variants.
6. genebass-gene-burden-skill
   • Retrieve rare-variant burden support for candidate genes.
7. clinvar-variation-skill (when include_clinvar=true)
   • Add variant clinical/coding annotations.
8. gnomad-graphql-skill (when include_gnomad_context=true)
   • Add frequency and gene-level constraint context.
9. human-protein-atlas-skill (when include_hpa_tissue_context=true)
   • Add tissue plausibility context for top genes.

Never perform additional retrieval after final candidate-gene scoring starts.

Output Contract (Required)

Always return:

1. locus_to_gene_mapping.json
2. locus_to_gene_summary.md

JSON contract

{
  "meta": {
    "trait_query": "...",
    "efo_id": "EFO_...",
    "generated_at": "ISO-8601",
    "sources_queried": []
  },
  "anchors": [
    {
      "rsid": "rs...",
      "grch38": {"chr": "3", "pos": 49629531, "ref": "A", "alt": "C"},
      "lead_trait": "...",
      "p_value": 2e-11,
      "cohort": "..."
    }
  ],
  "loci": [
    {
      "locus_id": "chr3:49000000-50200000",
      "lead_rsid": "rs...",
      "candidate_genes": [
        {
          "symbol": "MST1",
          "ensembl_id": "ENSG...",
          "overall_score": 0.71,
          "confidence": "High|Medium|Low|VeryLow",
          "evidence": {
            "l2g_max": 0.83,
            "coloc_max_h4": 0.84,
            "eqtl_tissues": ["Lung"],
            "rare_variant_support": "none|nominal|strong",
            "coding_support": "none|noncoding|coding",
            "clinvar_support": "none|present",
            "gnomad_context": "...",
            "hpa_tissue_support": ["lung"]
          },
          "rationale": [
            "..."
          ],
          "limitations": [
            "..."
          ]
        }
      ]
    }
  ],
  "cross_locus_ranked_genes": [
    {
      "symbol": "...",
      "supporting_loci": 3,
      "mean_score": 0.62,
      "max_score": 0.81
    }
  ],
  "warnings": [],
  "limitations": []
}

Markdown summary contract

The summary must include sections in this exact order:

1. Objective
2. Inputs and scope
3. Anchor variant summary
4. Per-locus top genes
5. Cross-locus prioritized genes
6. Key caveats
7. Recommended next analyses

Optional Figure Contract

Only produce figures when include_figures=true.

If figures are generated, append this block to JSON:

{
  "figures": [
    {
      "id": "locus_gene_heatmap",
      "path": "./output/figures/locus_gene_heatmap.png",
      "caption": "Top candidate genes by evidence component across loci"
    }
  ]
}

Recommended figure set:

1. locus_gene_heatmap.png
   • Rows: top genes, columns: evidence components (L2G, coloc, eQTL, burden, coding).
2. locus_score_decomposition.png
   • Stacked bars per locus for top 3 genes.
3. tissue_support_dotplot.png
   • Gene-by-tissue evidence dots from GTEx/HPA context.

If plotting dependencies are unavailable, skip PNG generation and output Mermaid diagrams in markdown as fallback. The script also returns inline_image_markdown and render_instructions fields to support inline chat rendering.

Scoring Rules (Deterministic)

For each candidate gene per locus, compute:

• l2g_component: max L2G score for the gene in locus (0..1)
• coloc_component: max h4 (or clpp when only CLPP is available), clipped to 0..1
• eqtl_component: min(1, relevant_tissue_hits / 3)
• burden_component:
  • 1.0 if burden p < 2.5e-6
  • 0.6 if 2.5e-6 <= p < 0.05
  • 0.0 otherwise
• coding_component:
  • 1.0 for coding consequence in target gene with supportive ClinVar annotation
  • 0.6 for coding consequence in target gene without supportive ClinVar annotation
  • 0.3 for noncoding-in-gene support only
  • 0.0 otherwise

Overall score:

overall_score = 0.40*l2g + 0.25*coloc + 0.15*eqtl + 0.10*burden + 0.10*coding

Confidence label:

• High if score >= 0.75
• Medium if 0.55 <= score < 0.75
• Low if 0.35 <= score < 0.55
• VeryLow if score < 0.35

Pipeline Contract

Phase 0: Validate and normalize input

• Enforce that at least one of trait_query, efo_id, seed_rsids is present.
• Normalize rsID formatting and deduplicate seed variants.
• Resolve free-text trait to one canonical EFO term when needed.

Phase 1: Build anchor set

• If trait/EFO input is provided, pull associations and rank anchors by p-value and effect availability.
• Merge trait-derived anchors with user-supplied seed_rsids.
• Cap anchors using max_loci and log dropped anchors in warnings.

Phase 2: Gather locus-to-gene evidence

• Normalize anchor coordinates (both builds when possible).
• Pull Open Targets locus evidence (credible set/L2G/coloc).
• Pull GTEx variant-level eQTL rows.
• Pull gene-level burden results for mapped candidate genes.
• Pull ClinVar and gnomAD context when enabled.

Phase 3: Harmonize and score

• Build a per-locus candidate-gene table.
• Compute deterministic component scores and overall score.
• Create cross-locus aggregate rankings.

Phase 4: Synthesize outputs

• Write JSON mapping file.
• Write markdown summary in exact section order.
• Optionally generate figures and append figures metadata.

Phase 5: QC gates

Fail the run when any of the following occurs:

• No anchors after normalization.
• Unresolved GRCh38 coordinates should be surfaced as status=degraded, not treated as an analytically clean pass.
• Any locus has candidate genes without score fields.
• overall_score outside 0..1.
• Summary section order mismatch.
• Claim of causality without explicit evidence support in rationale text.

Public Interface

def map_locus_to_gene(input_json: dict) -> dict:
    ...

Return:

{
  "status": "ok",
  "mapping_output_path": "./output/locus_to_gene_mapping.json",
  "summary_output_path": "./output/locus_to_gene_summary.md",
  "figure_paths": [],
  "warnings": [],
  "limitations": []
}

Non-Invention Rules

• Never invent rsIDs, p-values, scores, cohort labels, tissues, or gene links.
• Never silently impute missing evidence as positive support.
• When evidence is missing, record it as a limitation and reduce confidence.
• Keep evidence provenance explicit (source skill + endpoint family) in rationale lines.

Non-Goals

• Do not claim definitive causal genes from association evidence alone.
• Do not run fine-mapping methods not directly provided by upstream sources.
• Do not collapse multiple independent signals into one without stating assumptions.