Use AlphaGenome Atlas Without Treating AVI as a Diagnosis

Use AlphaGenome Atlas through the browser portal for visual, zero-code exploration or through the Python client when queries must be reproducible and filtered. In both routes, verify the genome assembly, chromosome, position and alleles before interpreting the AlphaGenome Variant Impact (AVI) score.

The result is a research-prioritization signal, not a diagnosis. AVI can move a variant higher or lower on an experimental shortlist, but it does not establish pathogenicity, disease causation or a patient-specific clinical conclusion.

Choose the portal or API by the job

The portal suits visual inspection and small exploratory queries; the Python route suits defined variant lists, recorded parameters and repeatable filters. Google’s Atlas launch description identifies the portal as a zero-code interface and describes AVI as a score that combines coding and non-coding predictions to prioritize research candidates.

Browser portal: choose it when you need to examine an individual variant and its predicted molecular context interactively.
Python Atlas client: choose it when you have a controlled input list, need consistent scorer or ontology filters, or must preserve outputs for later analysis.
Fresh AlphaGenome inference: use the separate model client only when precomputed Atlas records do not answer the question. Retrieving Atlas scores and generating new predictions are different operations.

Whichever path you choose, keep a query manifest containing the assembly, chromosome, position, reference allele, alternate allele and the biological question. This prevents a score from becoming detached from the precise variant and context that produced it.

Resolve coordinates before submitting anything

A variant position and analysis interval are reconciled to the same verified locus before an AlphaGenome Atlas query.

A coordinate mismatch can return a plausible-looking result for the wrong locus. The AlphaGenome API reference specifies that genome.Interval is zero-based and half-open, including start and excluding end, whereas genome.Variant uses a one-based, VCF-compatible position.

For a VCF row, pass POS to genome.Variant without subtracting one. When constructing an interval, convert its boundaries independently instead of reusing the one-based position as a zero-based start. Confirm that the input uses the human hg38 reference expected by AlphaGenome, retain the original REF and ALT alleles, and investigate any reference mismatch rather than changing an allele merely to make a query succeed.

Follow the zero-code portal path

Open the AlphaGenome Atlas portal and select its variant-query route.
Enter the supported variant identifier or genomic representation, preserving the original input in your manifest.
Before reading the score, compare the returned chromosome, position and alleles with the intended variant.
Use AVI to establish priority, then examine the molecular-effect views and biological contexts associated with that candidate.
Record the variant representation, relevant scorer or track context and the reason the result was retained or rejected.

Do not interpret AVI in isolation when your hypothesis depends on a particular gene, tissue, cell type or mechanism. An aggregate signal can prioritize a candidate while leaving the relevant mechanism unresolved; the supporting molecular outputs determine whether the prediction aligns with the system you can study.

Build a reproducible Python Atlas path

A Python Atlas client retrieves precomputed scores while preserving variant inputs, scorer metadata and reproducibility records.

Install the alphagenome package, obtain an API key and load the credential from an environment variable or secret manager. Import genome and the Atlas module, create the client with atlas.atlas.create, and represent each candidate as a genome.Variant using chromosome, one-based position, reference bases and alternate bases.

The official AlphaGenome repository identifies Atlas as a collection of precomputed variant-effect scores, AVI scores and feature importances; its Atlas client supports single-variant, variant-list and interval queries. Call scorer_metadata to discover available scorer names, then pass an explicit requested_scorers collection to query_variant, query_variants or query_interval. Ontology terms, gene identifiers and gene names can narrow results when those filters follow from the research question.

Save scorer metadata with the analysis so the requested names and contextual tracks remain interpretable.
Run one known candidate through query_variant to verify authentication, chromosome naming and the returned structure.
Use query_variants for a predetermined list; use query_interval only when the alternatives throughout a defined region are genuinely needed.
Persist the mapping from scorer names to returned AnnData objects together with package version, inputs, filters and failures.
Keep the API key outside shared notebooks and version-controlled files.

Read AVI and returned objects without overclaiming

AVI is a prioritization score. Its practical question is which variants deserve closer investigation, not which diagnosis a person has. A high rank can justify examining supporting molecular predictions, but it cannot by itself supply a clinical classification or prove a causal mechanism.

Variant identity: chromosome, one-based position, REF and ALT define the record being evaluated.
Scorer: identifies the molecular-effect calculation. Compare results only after confirming that the scorer and biological context are relevant.
Raw value: contains the scorer’s predicted effect value; scales from different scorer families are not automatically interchangeable.
Calibrated score or layer: places an output on the calibration supplied for that scorer. It is not a probability that a patient has a disease.
Track metadata: retains information such as assay, tissue, cell type or strand where applicable.
Feature importance: can suggest which sequence features contributed to a prediction, but does not experimentally establish the mechanism.

Direction must also be interpreted within the named scorer. A negative signed result is not automatically protective, and a large absolute value is not automatically pathogenic. Preserve the scorer name, sign convention, gene and contextual metadata whenever a result leaves the original object.

Match query scale to the rules and validation plan

An Atlas-prioritized variant moves to mechanism-specific experimental validation without being treated as a clinical diagnosis.

One or a few variants: use the portal or query_variant.
A defined candidate set: use query_variants, save completed results and retry failures without repeating successful work.
A bounded locus: use query_interval only when systematic retrieval across that interval serves the hypothesis; the client processes three single-nucleotide alternatives per position.
A large study: design for variable service capacity rather than assuming a fixed public quota. Atlas queries typically receive higher rates than fresh predictions, but rates vary with demand.

Under the repository’s stated terms, the API is free for non-commercial use, and most AlphaGenome outputs and Atlas information are limited to non-commercial use unless the applicable terms provide otherwise; commercial AlphaGenome access is available separately through Google Cloud. The same terms state that predictions are for theoretical modelling and research and must not be used for clinical decision-making or relied on for medical advice.

The handoff from prediction to evidence should be explicit. Use AVI to prioritize, identify the molecular output that supports the ranking, compare it with relevant population and functional annotations, and choose an experiment capable of testing the proposed mechanism. Depending on the hypothesis, that may be a reporter assay, targeted splicing assay, expression measurement or genome editing in a relevant cellular context.

Keep the Atlas result labelled as a model prediction throughout that process. Experimental evidence may support later interpretation under an appropriate research or clinical framework, but renaming a prediction as a classification does not change the evidence behind it.

Use AlphaGenome Atlas Without Mistaking a Prediction for a Diagnosis