Genomics

Setup

On first use, read setup.md for integration guidelines. Ask user consent before creating ~/genomics/ workspace.

When to Use

User has processed genomic data (VCF files) and needs clinical interpretation. Agent handles variant classification, pharmacogenomics recommendations, and annotation lookup. NOT for raw data processing — use bioinformatics skill for alignment and variant calling.

Architecture

Memory lives in ~/genomics/. See memory-template.md for structure.

~/genomics/
├── memory.md           # Context + preferences + interpretation history
└── cases/              # Active interpretation cases

Quick Reference

Core Rules

1. Classify Variants Using ACMG Guidelines

Every variant needs systematic classification:

Never classify without evidence. State "insufficient data" when appropriate.

2. Check Population Frequency First

Before clinical interpretation, verify frequency:

MAF >1% in any population = likely benign for rare disease.

3. Cross-Reference Multiple Databases

Single-source interpretation is insufficient. Triangulate evidence.

4. Report Pharmacogenomics Actionably

For drug-gene interactions, provide:

• Diplotype (e.g., CYP2D6 1/4)
• Predicted phenotype (poor/intermediate/normal/ultra-rapid metabolizer)
• Drug list affected
• Dosing guidance (CPIC/DPWG when available)

5. Separate Germline from Somatic Context

Always state which context you're interpreting.

6. Acknowledge Uncertainty

• Novel variants often lack evidence
• VUS ≠ benign — requires ongoing monitoring
• Reclassification happens (ClinVar updates monthly)
• Computational predictions are supportive, not definitive

Pharmacogenomics Reference

High-Priority Drug-Gene Pairs (CPIC Level A)

Phenotype Interpretation

Annotation Resources

Common Interpretation Traps

• Ignoring population specificity — Variants common in African populations may look rare in European-biased databases
• Trusting single ClinVar submitter — Check submitter count and review status (≥2 submitters, no conflict preferred)
• Conflating computational prediction with evidence — CADD/REVEL are supportive, not diagnostic
• Missing compound heterozygosity — Two VUS in trans can be pathogenic together
• Outdated database versions — gnomAD v4 has 800K+ exomes vs v2's 125K
• Ignoring gene-level constraint — pLI/LOEUF scores indicate tolerance to loss-of-function

External Endpoints

This skill does NOT automatically call external APIs. All database references are for manual lookup:

No automatic network requests. The skill provides URLs and guidance for manual lookup only.

Security & Privacy

Data that stays local:

• All interpretation work runs locally
• No variant data sent externally by this skill
• No automatic API calls to any database

This skill does NOT:

• Make network requests automatically
• Upload patient variants anywhere
• Connect to databases without explicit user action
• Store identifiable genomic information outside ~/genomics/

Related Skills

Install with clawhub install if user confirms:

• medicine — clinical decision support
• biology — molecular mechanisms
• chemistry — drug metabolism pathways
• health — patient care context

Feedback

• If useful: clawhub star genomics
• Stay updated: clawhub sync