Variant Analysis using AlphaGenome

Prerequisites

uv: Read the uv skill and follow its Setup instructions to ensure

uv is installed and on PATH.

User Notification: If LICENSE_NOTIFICATION.txt does not already exist in

this skill directory then (1) prominently notify the user to check the terms at https://deepmind.google.com/science/alphagenome/, then (2) create the file recording the notification text and timestamp.

.env file: Make sure the .env file exists in your home directory.

Create one if it does not exist.

ALPHAGENOME_API_KEY: This skill requires an API key to function.

You must ask the user for an API key if this skill looks relevant to their request and you do not have one in .env. The user can sign up at https://deepmind.google.com/science/alphagenome/. Do NOT ask the user to paste their key into the chat (this would leak the key into the agent's context). Instead, explain that a key is necessary to use AlphaGenome and give the user this command substituting ENV_FILE with the resolved literal path to the .env file:

    printf "Enter AlphaGenome API key (typing hidden): " && read -s key && echo && echo "ALPHAGENOME_API_KEY=$key" >> "ENV_FILE" && echo "Saved."

The scripts load credentials automatically via dotenv. NEVER read, print, or inspect the .env file or its variables (e.g. no cat, grep, echo, printenv, or os.environ.get on keys). Credentials must stay out of the agent's context.

When running in sandbox, dotenv.load_dotenv() will be a no-op, and instead the sandbox will read credentials and inject them directly.

Core Rules

NEVER run python3 or python3 -c directly. The system Python does not

necessarily have pandas, numpy, and other key dependencies. ALWAYS use uv run to run ALL Python code — including scripts, ad-hoc analysis files, and one-liners. Do not attempt to pip install or create new venvs — uv manages an isolated environment automatically.

Offline Only: NEVER use external APIs (e.g., MyGene.info, Ensembl REST)

for gene/transcript lookup. Use lookup_gene_info.py with the local GTF. If it fails, fix the environment/paths, do not switch to external APIs.

API Key is required: ALPHAGENOME_API_KEY must be set before running

any script (in sandbox, credentials are injected automatically).

Notification: If this skill is used, ensure this is mentioned in the

output.

Report Format: Always use the templates in docs/report-templates.md

for generating analysis reports, and ensure to include the table of top hits from the discovery scan.

Environment Setup & Troubleshooting

Python Environment

All scripts must be executed using uv run, which manages an isolated virtual environment with the correct dependencies via uv.

uv run <script_name> [args...]

For ad-hoc scripts (e.g., inline analysis code saved to a temp file), pass the full path instead of a short name:

uv run --project $SKILL_DIR /tmp/my_analysis.py --arg1 val1

[!NOTE] The first invocation resolves and installs dependencies (~10s). Subsequent runs use the cached environment and start instantly. The cache lives in ~/.cache/uv/.

Common Issues

Column Names: tidy_scores and metadata often use gene_name (not

gene_symbol) and output_type (not modality). Always inspect df.columns before filtering.

Large Genes: Genes > 500kb (e.g., USH2A) break the whole_gene view.

Use --view detail or manual regional windows instead.

Sashimi Strand Error: plot_components.Sashimi does NOT accept a

strand argument directly. Filter input tracks instead.

KeyError: 'ontology_curie': Not all tracks have ontology_curie. Check

track.metadata.columns before filtering.

Python Path: If exec: "python": executable file not found occurs,

ensure you are using uv run instead of bare python/python3.

NotImplementedError (pandas): "iLocation based boolean indexing on an

integer type is not available". This occurs when using boolean masks with .iloc on integer-indexed DataFrames in newer pandas versions. Fix: Convert boolean masks to integer indices using np.flatnonzero(mask).

GTF Feather Case Sensitivity: The AlphaGenome GTF Feather file uses

Capitalized column names (Feature, Start, End, Strand) unlike standard GTF files. Always check df.columns if getting KeyErrors.

score_variant ontology filtering: score_variant does NOT accept

ontology_terms as an argument. You must filter the returned AnnData objects manually by inspecting adata.var columns. In contrast, predict_variant DOES accept ontology_terms directly.

Sashimi Zoom Logic: To ensure "skipping" arcs are visible, expand the

zoom to include the flanking exons rather than relying on junction overlap alone.

Junction Scores: Raw Junction objects from prediction may be simple

Intervals. Use junction_data.get_junctions_to_plot(predictions=..., name=...) to retrieve objects with the .k (abundance/score) attribute.

uv Not Found: If exec: uv: not found, follow the installation

instructions in Prerequisites.

Registry Authentication Error (401): If uv fails with 401 Unauthorized

for a private registry, set UV_INDEX_URL=https://pypi.org/simple before running the script.

References

alphagenome-api.md — API reference and code

patterns

interpretation-guide.md — Interpretation

guide, score magnitude rules, ISM, and checklist.

report-templates.md — Full report templates
scripts/visualize_variant_effects.py

— Single-variant visualization template (Ref/Alt comparisons, Splicing).

Splicing Zoom Strategy: Uses a Hybrid Approach for optimal

visibility:

Base Interval: Variant +/- 1 downstream and upstream exon

(Structural Context).

Junction Expansion: Expands to include the full span of any

significant splicing junction (e.g., exon skipping events that span multiple exons).

Anchor Enforcement: Ensures the exons anchoring these long

junctions are fully visible. Lesson: Simple fixed windows (e.g., 2kb) or nearest-exon logic often fail for skipping events. Always use the observed junction data to drive zoom levels.

examples/splicing/ — Splicing analysis examples
examples/model_limitation_RNU4ATAC/

— ncRNA structure limitation case study

examples/polyadenylation_HBA2/ — 3'

UTR / Polyadenylation case study

examples/regulatory/ — Regulatory variant

examples

examples/negative_result_GATA4/ —

Negative results (mathematical artefact)

examples/negative_result_TGFB3/ —

Negative results (proxies)

scripts/lookup_gene_info.py — Gene &

transcript lookup

scripts/resolve_ontology_terms.py —

Ontology term resolution (UBERON/CL IDs)

--------------------------------------------------------------------------------

Code Patterns

Broad Discovery Scan

Use score_variant across differential scorers only to discover unexpected tissue effects.

from alphagenome.models import dna_client
from alphagenome.models import variant_scorers
from alphagenome.data import genome
import os
import pandas as pd

# Setup API Key and Client
dna_model = dna_client.create(api_key=os.environ.get('ALPHAGENOME_API_KEY'),
                              address='dns:///gdmscience.googleapis.com:443')

# Define Variant (example)
variant_str = "chr2:1234:A>C"
chrom, pos_str, ref_alt = variant_str.split(':')
ref, alt = ref_alt.split('>')
pos = int(pos_str)

# Use supported sequence length (e.g., 2**20 for optimal performance)
SEQ_LENGTH = 2**20
interval = genome.Interval(chrom, pos - SEQ_LENGTH // 2, pos + SEQ_LENGTH // 2)
variant = genome.Variant(chrom, pos, ref, alt)

scorers = [
    variant_scorers.RECOMMENDED_VARIANT_SCORERS[m]
    for m in variant_scorers.RECOMMENDED_VARIANT_SCORERS
    if "ACTIVE" not in m and "CAGE" not in m and "PROCAP" not in m
]

print(f"Scoring variant {variant_str}...")
scores_list = dna_model.score_variant(interval=interval, variant=variant, variant_scorers=scorers)

# Process and Display Results
all_dfs = []
for score_adata in scores_list:
    df = variant_scorers.tidy_scores([score_adata], match_gene_strand=True)
    if df is not None:
        all_dfs.append(df)

if all_dfs:
    df = pd.concat(all_dfs)
    significant = df[df['quantile_score'].abs() > 0.995]
    ranked = significant.sort_values('raw_score', key=abs, ascending=False)
    print("Top Significant Hits:")
    print(ranked[['biosample_name', 'gene_name', 'output_type', 'quantile_score', 'raw_score']])

Extended Search for Disease-Relevant Tissues

# Define keywords based on disease context
disease_keywords = ["liver", "hepatocyte"]

# Filter for any match
mask = df['biosample_name'].str.contains('|'.join(disease_keywords), case=False, na=False)

relevant_hits = df[mask].sort_values('raw_score', key=abs, ascending=False)
print(f"\n--- Extended Analysis (Keywords: {disease_keywords}) ---")
print(relevant_hits.head(20)[['biosample_name', 'output_type', 'raw_score', 'quantile_score']])

Workflow Checklist

Variant Analysis Progress:
- [ ] Step 0: Review Golden Examples (MANDATORY)
- [ ] Step 1: Create Output Folder and Setup
- [ ] Step 2: Parse User Query & Research
- [ ] Step 3: Resolve Tissues & Modalities
- [ ] Step 4: Visualize & Save Plots
- [ ] Step 5: Analyze Predictions (view plots, no code). MANDATORY: Read [interpretation-guide.md](docs/interpretation-guide.md) before interpreting results.
- [ ] Step 6: Write Report, save it as `report.md` (MANDATORY)
- [ ] Step 7: Self-Critique (view `report.md` to verify links & claims)
- [ ] Step 8: Make artifact out of `report.md`