Genomic Variant Research Skill

This repository turns a real variant-classification workflow into a reusable, auditable skill for genetic variant analysis. A specialist should review every output.

An AI agent skill for systematic computational analysis of human genetic variants. Feed it a gene + variant → get a full ACMG evidence package using 30+ free bioinformatics tools.

Built for Claude Code and compatible AI agents.

What It Does

Input: Gene + Variant (e.g., STRC c.4976A>C)
    │
    ├─ Step 1: Coordinates & annotation (Ensembl VEP)
    ├─ Step 2: Pathogenicity scores (AlphaMissense + REVEL + CADD)
    ├─ Step 3: Population frequency (gnomAD + ClinVar)
    ├─ Step 4: Protein structure (AlphaFold DB + AF3)
    ├─ Step 5: Splicing & regulatory (SpliceAI + AlphaGenome)
    ├─ Step 6: Conservation (UniProt orthologs + Clustal Omega + OrthoDB)
    ├─ Step 7: Disorder & domains (IUPred3 + ESMFold)
    ├─ Step 8: Clinical context (ClinGen + OMIM + ClinicalTrials.gov)
    │
Output: ACMG evidence summary with classification

Tools Covered (36)

Category	Tools
Structural Biology	AlphaFold DB, AlphaFold 3, AlphaGenome, SWISS-MODEL, ESMFold, IUPred3, Mol*, PyMOL
Variant Scoring	AlphaMissense, REVEL, CADD, SpliceAI, dbNSFP, VarSome, InterVar, Franklin
Population Genetics	gnomAD, ClinVar, seqr, DECIPHER, ClinGen
Conservation	UniProt, UCSC Genome Browser, Ensembl REST API, OrthoDB, Clustal Omega
Disease-Specific	OMIM, Deafness Variation Database, Hereditary Hearing Loss Homepage
Gene Therapy	ClinicalTrials.gov, AAV Capsid Database, ASGCT
Protein Analysis	SignalP, NetGPI, NetNGlyc
Literature	PubMed/PMC

All tools are free. No wet lab required. Total compute cost: ~$50-100 in LLM API calls.

Install

Copy SKILL.md into your agent's skills directory:

# For Claude Code / Hermes
cp SKILL.md ~/.claude/skills/genomic-variant-research/SKILL.md

# Or clone the whole repo
git clone https://github.com/lyfar/genomic-variant-research.git ~/.claude/skills/genomic-variant-research

Then ask your agent: "Score variant BRCA1 c.5266dupC using the genomic variant research workflow"

Hard-Won Pitfalls

These cost us weeks to discover. They're all in the skill, but the critical ones:

1. OrthoDB sequences are UNALIGNED — never check positional conservation by taking position N from raw FASTA. Sequences have indels. Must pairwise-align a window around the target. We got "3.8% conservation" for a 100%-conserved position before catching this.

2. PP1 ≠ conservation — PP1 (co-segregation) requires affected family members, not cross-species conservation. Conservation data supports PP3. Mixing these up changes classification.

3. SIFT/PolyPhen fail for genes with pseudogenes — pseudogene sequences contaminate the alignment, making conserved positions appear variable. Use AlphaMissense and REVEL instead.

4. Don't double-count PP3 — AlphaMissense and REVEL measure overlapping signal. Use the strongest single predictor for the PP3 evidence level.

5. CADD direct API returns "Not found" — use Ensembl VEP with ?CADD=1 instead.

6. gnomAD constraint: low pLI is EXPECTED for recessive genes — check missense Z-score instead.

Background

This skill was developed during a real variant classification project for STRC-related hearing loss (DFNB16). The worked example in the skill (STRC E1659A → VUS-high) is a real case that led to contact with researchers at Harvard Medical School and Fudan University.

The full research is published at strc.egor.lol.

License

MIT. Use it, fork it, improve it. If you find a new pitfall, please open an issue or PR.