AlphaGenome 2 Predicts Gene Changes by Cell Type

AlphaGenome 2 predicts how genetic variants affect specific cell types, with reported gains over a fine-tuned AlphaGenome 1 benchmark.

Quick Answer

AlphaGenome 2 predicts how genetic variants affect gene activity in a specified human cell type, giving researchers a more targeted way to investigate disease mechanisms and drug responses. Google Deepmind reports that the model identified about 30% of causal variants in the TenK10K benchmark, compared with 16% for a fine-tuned AlphaGenome 1.

Evidence and Access Status

  • Official announcement, November 12, 2026: Google Deepmind’s announcement is the source for the model’s capabilities, training scale, reported benchmark results, access terms and planned product integrations.
  • Research context, June 3, 2024: The Context-Specific Variant Prediction article in Frontiers in Systems Biology describes the challenge of determining the functional effects of noncoding variants.
  • Independent validation status: The reported TenK10K, vaccine-response and inflammatory bowel disease results are presented in Google Deepmind’s materials. The available material does not identify these reported results as independently reproduced outside the company’s evaluation.
  • Access: AlphaGenome 2 is available for noncommercial research through an API, while commercial users have access through a closed preview. It is not approved for clinical diagnostics.

What Is New in AlphaGenome 2?

AlphaGenome 2 evaluates a DNA sequence against one specified cell state rather than spreading predictions across many tissues at once. It predicts gene expression, DNA accessibility and DNA packaging within the nucleus, and it also predicts DNA methylation, the chemical tags that can activate or silence genes. The model was trained on more than 107 million individual cell profiles drawn from 68 human and mouse tissues. It can also infer missing properties from a supplied measurement, such as estimating DNA packaging when researchers provide gene-activity data. Google Deepmind says these predictions can help researchers choose laboratory experiments to run next.

Capability and Limitation Table

Area What AlphaGenome 2 Does Status or Limitation
Cell-specific prediction Evaluates DNA sequences in a specified cell state. A cell state means the biological condition of a cell, such as resting or activated.
Gene expression Predicts the strength of gene activity and whether a variant increases or decreases it.
DNA structure Predicts DNA packaging and which genome regions are accessible. These are computational predictions, not clinical measurements.
DNA methylation Predicts chemical tags that can activate or silence genes. Reported as a new capability of AlphaGenome 2.
Unseen cell states Can calculate effects using gene-activity data from a few dozen cells without retraining the full model. Google Deepmind reports that it often outperforms specialized models trained directly on the same data.
Clinical use Supports genomic and drug-development research. The system is not approved for clinical diagnostics.

How Does Cell-Type Context Change Prediction?

Human cells carry the same DNA blueprint, but different cell types switch different genes on or off. AlphaGenome’s predecessor accepted a DNA sequence as long as one million letters and predicted properties across thousands of experiments spanning many tissues simultaneously. Google Deepmind said this approach spread the model’s capacity across tissues and made subtle differences between closely related cell types harder to capture.

AlphaGenome 2 instead receives the DNA sequence and a specific biological state as inputs. In the COVID-19 mRNA vaccine dataset, it separated vaccine-activated T cells from resting T cells. It also found that a variant associated with severe COVID-19 disrupted activity mainly in activated cells, an effect that was not visible in resting cells.

What Do the Benchmark Results Show?

The following results are Google Deepmind-reported. The fine-tuned baseline is AlphaGenome 1 adjusted for the particular benchmark rather than the untouched original model.

Metric AlphaGenome 2 Fine-Tuned AlphaGenome 1 Evidence Status
Causal-variant identification on TenK10K About 30% 16% Google Deepmind-reported
Correct prediction of expression direction 43% 23% Google Deepmind-reported
TenK10K cell coverage 28 primary immune cell types Same benchmark Google Deepmind-reported
Blood stem-cell maturation Predictions closer to observed measurements Less accurate predictions Google Deepmind-reported comparison

How Does AlphaGenome 2 Reproduce a Disease Mechanism?

In a computational reproduction of an inflammatory bowel disease mechanism, AlphaGenome 2 isolated a DNA variant in a regulatory switch located far from the affected gene. The variant selectively increased ETS2 activity in gut macrophages while other cell types and genes remained largely unaffected. This was a computational reproduction of a mechanism reported in earlier research, not a clinical test of AlphaGenome 2.

How Can AlphaGenome 2 Support Genome-Wide Studies?

Genome-Wide Association Studies connect broad DNA regions with traits such as height or type 2 diabetes, but often locate signals only roughly. Google Deepmind has precomputed predictions for hundreds of cell and tissue conditions. The model can narrow the candidate causal variants, identify affected cell types such as brain astrocytes or specific immune cells, and predict whether a mutation increases or decreases gene expression when its confidence is high enough.

The same predictions are intended partly to support AI-assisted drug development. For cell-specific adaptation, researchers can supply gene-activity data from a few dozen cells, and the model calculates effects for that previously unseen state without retraining the full system.

Context-Specific Variant Prediction (2024)

Context-Specific Variant Prediction, published in Frontiers in Systems Biology on June 3, 2024, addresses the prediction of functional effects for noncoding genetic variants from DNA sequences. Its research abstract says conventional Genome-Wide Association Studies, Transcriptome-Wide Association Studies and Quantitative Trait Loci studies are constrained by obscured molecular-level mechanisms. Cell-specific prediction addresses that gap by considering the biological context in which a variant may act.

What Do Cell-Specific Genomics Studies Add?

Practical Checklist for Evaluating a Cell-Specific Prediction

  • Verify the intended use and access terms.
  • Identify the cell state and tissue represented.
  • Compare the output with existing evidence.
  • Check whether the result has independent validation.
  • Record the model and benchmark versions.
  • Obtain qualified expert review before making clinical or commercial decisions.

What to Do Now

  1. Review the noncommercial API terms and closed-preview access criteria.
  2. Record the intended cell state, tissue and biological condition.
  3. Compare predicted variants and expression changes with available experimental evidence.
  4. Identify whether each benchmark result is company-reported or independently reproduced.
  5. Consult a qualified genetics specialist before using results in clinical or commercial decisions.

Key Considerations

  • The reported benchmark results come from Google Deepmind’s evaluation of AlphaGenome 2 and a fine-tuned AlphaGenome 1.
  • The available material does not identify the company’s reported benchmark results as independently reproduced outside its evaluation.
  • The training data includes more than 107 million individual cell profiles spanning 68 human and mouse tissues.
  • The model evaluates a specified cell state rather than producing one prediction across every tissue.
  • Noncommercial research access is available through an API, while commercial access is limited to a closed preview.
  • AlphaGenome 2 is not approved for clinical diagnostics.

Glossary

Cell state
The biological condition of a cell, such as resting or activated.
Causal variant
A genetic variant identified by the model as affecting a measured biological process.
Fine-tuned baseline
AlphaGenome 1 adjusted for the particular benchmark being tested.
Upregulation
Increased gene activity.
Noncoding variant
A genetic variant outside the protein-coding regions of a gene.
Noncommercial research
Research use that does not involve commercial activity, as defined by the applicable access terms.
Closed preview
Restricted access for commercial users before general commercial availability.

FAQ

Who Can Access AlphaGenome 2?

AlphaGenome 2 is available for noncommercial research through an API, while commercial users have access through a closed preview.

Is AlphaGenome 2 Approved for Clinical Diagnostics?

No. Google Deepmind says AlphaGenome 2 is not approved for clinical diagnostics.

What Does Identifying 30% of Causal Variants Mean?

On the TenK10K benchmark, AlphaGenome 2 identified about 30% of causal variants across 28 primary immune cell types.

How Does AlphaGenome 2 Compare With AlphaGenome 1?

On TenK10K, Google Deepmind reports about 30% causal-variant identification for AlphaGenome 2 versus 16% for fine-tuned AlphaGenome 1. In 43% of cases, AlphaGenome 2 also correctly predicted whether a variant upregulates or downregulates a gene. A fine-tuned version of AlphaGenome 1 managed just 16 and 23 percent.

Can AlphaGenome 2 Analyze a New Cell State?

Google Deepmind says researchers can supply gene-activity data from a few dozen cells and that AlphaGenome 2 can calculate effects for a previously unseen cell state without retraining the full model.

This article reports on publicly available information from the sources listed below. We have no reason to believe it is inaccurate, but we have not independently verified every claim, and details may change as the story develops. This is not financial, legal or investment advice.

Sources


See our editorial disclaimer for how our articles are produced.

🤖
Is your business visible to AI assistants?

Run a free scan to see your AI Visibility Score, SEO rating, and local citation accuracy.

Check Your Score →