Predict the regulatory effect of a non-coding variant

Hand Claude Code a single chr:pos:ref>alt variant and get back AlphaGenome’s predicted effect on expression, chromatin accessibility, histone marks, splicing, and transcription-factor binding — with tissue-resolved tracks and an in-silico mutagenesis logo for the disrupted motif.

   
Problem class Knowledge synthesis
Subject areas Molecular and Cellular Biology, Integrative Structural and Computational Biology
Evidence level Reported
Complexity One skill or MCP
Availability Subscription required
Compute Laptop

Problem

A GWAS hit or a rare-disease exome lands on a variant that is nowhere near a coding region. ClinVar is silent, the variant is in an intron or an intergenic enhancer, and the question is mechanistic: does this base change touch a regulatory element, and if so, which gene, which tissue, and which molecular readout (expression up/down, a lost TF footprint, a cryptic splice site)? Wet-lab follow-up — a reporter assay, an eQTL look-up — is slow and not always available for the right cell type. A sequence-to-function model can generate the testable hypothesis in seconds. Solved looks like: a ranked summary of the variant’s predicted molecular consequences across relevant tissues, the affected gene, and a sequence logo showing what regulatory motif the alt allele breaks or creates.

  1. Get an AlphaGenome API key. Sign up at deepmind.google.com/science/alphagenome, accept the non-commercial research terms, and copy the key. Install uv if you don’t have it:

    curl -LsSf https://astral.sh/uv/install.sh | sh
echo "ALPHAGENOME_API_KEY=<your-key>" >> ~/.env

  2. Install the AlphaGenome single-variant skill from DeepMind’s science-skills collection (manual copy is the Claude path):

    git clone https://github.com/google-deepmind/science-skills
cp -r science-skills/skills/alphagenome_single_variant_analysis ~/.claude/skills/
cp -r science-skills/skills/scienceskillscommon ~/.claude/skills/

  3. Resolve the tissue context first. The model is tissue-resolved, so name the cell type that matters for your phenotype. Have Claude run the skill’s resolve_ontology_terms.py to map “pancreatic beta cell” or “CD4 T cell” to its UBERON/CL ID before scoring.

  4. Score the variant. A minimal prompt:

    Run the alphagenome_single_variant_analysis skill on
chr11:5226774:T>A (GRCh38). Score RNA-seq expression, DNASE
accessibility, ChIP histone marks, and TF binding in the
relevant blood/erythroid cell types. Identify the affected
gene, rank the modalities by effect size, run analyze_ism.py
for the disrupted motif, and run interpret_splicing.py if any
splice signal changes.

  5. Read the ranked output and the ISM logo. The skill emits reference-vs-alternate tracks per modality, a splicing-disruption analysis, and an in-silico mutagenesis sequence logo. Treat the top-ranked modality as the lead hypothesis (e.g., “abolishes a GATA1 footprint → reduces enhancer accessibility → lowers target expression in erythroid cells”) and confirm the gene assignment with the skill’s offline GTF lookup before reporting.

Why this assembly

Rung 2 of the simplicity ladder. The hard part — a 1 Mb sequence-to-function model spanning every regulatory modality — is the AlphaGenome API; the skill is the thin orchestration layer that resolves ontologies, calls the API, and renders tracks and ISM logos. Plain Claude Code (rung 1) cannot predict molecular phenotypes from sequence; it has no model. There is nothing to escalate to at rung 3/4 — this is a single-model query, and the model is state of the art for the task. The one judgment call the recipe adds is naming the right tissue, because cell-type-specific regulation is exactly where the model is weakest.

Availability

Subscription required (free-tier, but gated). The AlphaGenome API is a signup-gated research preview, free for non-commercial use with an API key and acceptance of the terms — commercial use is not covered. The skill code itself is OSS (Apache-2.0 code, CC-BY-4.0 docs). Note: the skill’s primary npx skills add install path targets Gemini/Antigravity; for Claude use the manual copy shown above.

Compute requirements

Laptop-sufficient on the client side — all heavy computation runs server-side on the AlphaGenome API. A single-variant scoring call returns in roughly a second to a few seconds of model time; the skill installs its own Python deps via uv on first run. No local GPU. Network access to the API is required.

Evidence

Reported. AlphaGenome is peer-reviewed: Advancing regulatory variant effect prediction with AlphaGenome, Nature 2026 reports that the model matches or exceeds the strongest available external models on 25 of 26 variant-effect-prediction evaluations, and on 22 of 24 single-sequence prediction tasks — while being the only model that jointly predicts all assessed modalities. The paper recapitulates the mechanisms of clinically relevant variants near the TAL1 oncogene as a worked case. Experts note a caveat the recipe inherits: cell-type-specific regulation remains the model’s weakest dimension, so tissue choice matters.

No peer-reviewed benchmark of “Claude + the AlphaGenome skill” specifically is known; the agent loop wraps the published API and adds ontology resolution and visualization, not a new model. The quantitative anchor is the Nature paper.

Alternatives considered

  • Interpret a clinical variant. Reach for that recipe when the variant is coding or already in ClinVar/gnomAD — it does database-anchored clinical interpretation. This AlphaGenome recipe is for the non-coding, regulatory, mechanism-unknown case where the databases are silent.
  • eQTL/regulatory database lookup (GTEx). If your variant is a known eQTL in the tissue you care about, a GTEx lookup is cheaper and observational rather than predicted. AlphaGenome wins when no eQTL exists for your variant or your cell type.
  • Plain Claude Code. Cannot predict sequence-to-function effects; it would only summarize what databases already say. Not an option for novel non-coding variants.

See also

Sources

Tried this recipe?

Share feedback — what worked, what didn’t, what you’d change. The form opens with this recipe pre-selected and a link back to this page.