AlphaGenome Maps 9 Billion Variants—Its Score Is a Filter, Not a Diagnosis

In its September 8, 2026 release of AlphaGenome Atlas, Google DeepMind made precomputed molecular-effect predictions for 9 billion possible single-nucleotide variants searchable through a web portal and accessible through the AlphaGenome API. The release also introduced the AlphaGenome Variant Impact score, or AVI, to help researchers rank variants for further investigation.

The boundary is crucial: Atlas entries are model outputs, not observations from a patient or laboratory proof of disease. Nature’s independent launch coverage described the resource as freely available for non-commercial use while stressing that it cannot replace experiments or the individual evidence needed when diagnosing disease.

What AlphaGenome Atlas contains

The Atlas is comprehensive in the substitutions it enumerates, not in experimentally established consequences. Each record asks what AlphaGenome predicts would happen if one letter in the reference human genome were replaced by another; it does not show what has been measured in a person carrying that change.

The underlying model estimates effects on molecular processes including gene expression, RNA splicing and chromatin activity across biological contexts. That scope is particularly relevant outside protein-coding regions, where variants may change when, where or how strongly a gene operates without altering the protein sequence itself.

AVI is a summary layer over those predictions. It combines regulatory predictions from AlphaGenome with protein-impact information from AlphaMissense, producing one rankable value for coding and non-coding variants. Feature attributions provide a route back to the proposed mechanism by indicating whether predicted changes in splicing, expression, chromatin accessibility or protein function contributed to the score.

A record therefore contains two distinct kinds of output: a compact value for sorting candidates and more detailed predictions for forming a biological hypothesis. The first can determine which record a researcher examines sooner; neither establishes that the variant causes a phenotype.

From variant lookup to laboratory validation

A defensible workflow begins with a defined phenotype, locus or candidate set. The Atlas removes the need to run AlphaGenome separately for every substitution, but it cannot decide which tissue, gene or experimental system is relevant to the research question.

1. Look up a variant or region in the portal, or retrieve records programmatically through the API. Treat the AVI value and all molecular effects as predictions.
2. Rank candidates within a biologically coherent study set. Comparing unrelated variants solely by their top-line scores can hide differences in mechanism and context.
3. Inspect the feature contributions behind a high-ranking result. A predicted splice disruption calls for a different test from a predicted change in regulatory activity or protein function.
4. Compare the model output with independent evidence, such as phenotype fit, inheritance, population frequency and established gene–disease relationships when those categories apply.
5. Choose an assay that measures the proposed mechanism in an appropriate system, then record the observed result separately from the original prediction.

This workflow preserves three different claims: the model forecasts a molecular effect, AVI helps prioritize that forecast, and an experiment tests whether the effect occurs. Even a positive functional result does not by itself prove that a variant caused a patient’s condition; clinical interpretation depends on the wider evidence chain.

Why a single AVI score can mislead

A one-number score is useful because it discards detail, and risky for exactly the same reason. Similar AVI values can represent different predicted mechanisms, tissues or molecular consequences. A universal cutoff could therefore elevate a technically strong but biologically irrelevant signal over a candidate that better fits the study.

Context also has a physical limit. IEEE Spectrum’s assessment of the model’s 1 million-base-pair field of view notes that some enhancers regulate genes across still greater distances, while complex diseases can involve multiple variants whose combined effects are not captured by one Atlas entry.

Those limits matter at both ends of a ranked list. A high AVI value does not prove pathogenicity, and a low value does not prove harmlessness. Long-range regulation, interactions among variants, indirect effects or an inadequately represented biological context can weaken the connection between a predicted molecular change and an actual phenotype.

Molecular impact is not equivalent to clinical significance. A variant may measurably alter expression or splicing without causing disease, while an important disease mechanism may be one the model handles poorly. AVI can narrow an experimental queue; it cannot independently classify a patient’s variant, establish causality or produce a diagnosis.

Access has widened, but the evidence standard has not changed

The browser portal lowers the entry barrier for researchers who do not write code, while the API supports programmatic queries and larger workflows. Atlas access is currently framed for non-commercial research; commercial access through Google Cloud was described at launch as forthcoming.

The release changes the beginning of variant research by making predictions available before an experiment is chosen. The later stages remain intact: researchers must check whether the proposed mechanism fits other genetic and biological evidence, select a suitable system and measure what actually happens.

AlphaGenome Atlas is therefore a large hypothesis map rather than a catalogue of diagnoses. Its value will depend on whether ranked candidates survive independent benchmarking, mechanism-specific laboratory tests and, when patient care is involved, established clinical interpretation procedures.

Also read:

• Tongue AI’s 98.71% Score Was Color Accuracy, Not Diagnosis
• China’s First Moon Landing Does Not Depend on Its Reusable Super-Heavy Rocket
• Use Hashtags for Context, Not Reach: Instagram Now Caps Them at Five