uv: Read the uv skill and follow its Setup instructions to ensure
uv is installed and on PATH.
User Notification: If
.licenses/alphagenome_single_variant_analysis_LICENSE.txt does not already
exist in the workspace root 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 can register for a key at https://deepmind.google.com/science/alphagenome/.
You MUST use the safe credentials protocol in the credentials skill to
check for and request this key if this skill looks relevant to the user's request.
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.
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.
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.
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
import dotenv
# Load environment variables from ~/.env
dotenv.load_dotenv(os.path.expanduser('~/.env'))
# 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`
Multi-Variant Workflow
If multiple variants are specified, spawn sub-agents to run each variant
analysis and then synthesize each report.md into a single report.
