alphagenome-variant-analysis

name	alphagenome-variant-analysis
description	Analyze how a single non-coding genetic variant affects gene expression, chromatin accessibility, histone marks, transcription factors, and splicing using the AlphaGenome API. Use when the user gives a variant in chr:pos:ref/alt form and asks about regulatory or splicing effects, tissue-specific impact, or pathogenicity of a non-coding change. Requires an AlphaGenome API key.
license	Apache-2.0

AlphaGenome Single-Variant Analysis (Codex-native)

Codex-native adaptation of Google DeepMind's alphagenome-single-variant-analysis skill. It keeps the upstream workflow and interpretation rules but maps execution onto Codex tools instead of a bundled script harness.

Attribution: adapted from google-deepmind/science-skills (Apache-2.0 + CC-BY-4.0). Model terms: https://deepmind.google.com/science/alphagenome/

Gates (ask before proceeding)

• Credential: this workflow calls the AlphaGenome API and needs ALPHAGENOME_API_KEY. Do not read, print, or hardcode the key. Confirm the user has it exported (e.g. in ~/.env) before running anything. If absent, stop and ask them to register at the model terms URL above.
• Terms notice: on first use in a workspace, tell the user to review the AlphaGenome terms, and record that you did so in the run's provenance.
• Package install / network: the API client (alphagenome, pandas, numpy) is installed on demand. Ask before installing, and run all Python through uv run — never bare python3, never pip install into the system environment.

Workflow

1. Parse the variant. Accept chr:pos:ref/alt (or ref>alt). Split into chrom, pos (int), ref, alt. Reject malformed input.

2. Set up an output folder for plots, the JSON of scores, and report.md.

3. Broad discovery scan. Score the variant across the recommended differential scorers to find unexpected tissue effects, then rank by absolute effect. Run as a uv run script:

   import os, dotenv, pandas as pd
   from alphagenome.models import dna_client, variant_scorers
   from alphagenome.data import genome
   
   dotenv.load_dotenv(os.path.expanduser("~/.env"))
   model = dna_client.create(api_key=os.environ["ALPHAGENOME_API_KEY"],
                             address="dns:///gdmscience.googleapis.com:443")
   
   chrom, pos_s, ref_alt = "chr2:1234:A>C".replace("/", ">").split(":")
   ref, alt = ref_alt.split(">"); pos = int(pos_s)
   L = 2**20
   interval = genome.Interval(chrom, pos - L // 2, pos + L // 2)
   variant = genome.Variant(chrom, pos, ref, alt)
   scorers = [variant_scorers.RECOMMENDED_VARIANT_SCORERS[m]
              for m in variant_scorers.RECOMMENDED_VARIANT_SCORERS
              if all(t not in m for t in ("ACTIVE", "CAGE", "PROCAP"))]
   dfs = [variant_scorers.tidy_scores([a], match_gene_strand=True)
          for a in model.score_variant(interval=interval, variant=variant,
                                        variant_scorers=scorers)]
   df = pd.concat([d  d  dfs  d   ])
   top = df[df[].() > ].sort_values(, key=, ascending=)
   (top[[, , , , ]])
   

Boundaries

• Predictions are computational hypotheses, not clinical or diagnostic conclusions. Do not assert pathogenicity, disease causation, or treatment implications from these scores alone.
• Do not substitute external gene/annotation APIs for the model's own outputs when reproducing the upstream method.
• Preserve failures, uncertainty, and the model's known limitations (e.g. weak performance on some ncRNA structures).