- 5 min read
- 7 key claims
- 1 source cited
- For Developers, Businesses, AI enthusiasts
Overview
Google DeepMind announced AlphaGenome 2 as a next-generation model for predicting how DNA sequences and variants affect particular human cell types. The company describes the model as a tool for narrowing variant interpretation from broad tissue averages to individual cell types and states [vendor_claim] [1]. A shorter excerpt of the same announcement repeats this central description [vendor_claim] [1].
The model is available for non-commercial research through the AlphaGenome API, with a private preview for commercial research on Google Cloud [vendor_claim] [1]. The announcement does not state API limits, commercial-preview eligibility criteria, or pricing [1].
Why it matters
Cell-type-specific variant prediction could help researchers prioritize variants that alter gene activity only in particular cells, including cells relevant to disease [hypothesis] [1]. The model's cross-modality imputation feature may also help teams decide which experiments to run next by estimating an unmeasured property from one observed measurement [hypothesis] [1].
These are potential research workflows based on the model's stated capabilities. Their value in practice would depend on independent validation, which the announcement does not provide [1].
Opportunities
Possible applications include:
- Integrating cell-type-specific variant scores into genomics research pipelines [hypothesis] [1].
- Adding cross-modality imputation to single-cell analysis workflows to estimate missing measurements [hypothesis] [1].
- Prioritizing drug targets by examining predicted variant effects in disease-relevant cell types [hypothesis] [1].
- Adapting the model to newly profiled cell states without retraining, as the company says is possible [vendor_claim] [1].
Risks
All reported performance figures are vendor claims from a single TenK10K comparison and have not been validated by an independent third-party evaluation [1]. The announcement does not provide the benchmark protocol, uncertainty estimates, statistical significance values, or separate scores for each comparator, so the scope of the reported gains is unclear [1].
Errors in variant-effect predictions could misdirect research prioritization or downstream analysis [hypothesis] [1]. Commercial availability is limited to a private preview on Google Cloud, creating uncertainty about access terms, cost, and potential platform dependence [hypothesis] [1]. The evidence does not address how patient data would be handled or what regulatory requirements would apply to clinical uses [1].
Context
The first AlphaGenome model predicted the effects of DNA variants on molecular activity across broad tissues. Its companion, AlphaGenome Atlas, contains predictions for all nine billion possible single-letter changes in the human genome, according to Google DeepMind [vendor_claim] [1].
Those tools work from tissue-level averages, which can obscure differences between closely related cell types. AlphaGenome 2 is intended to address that limitation by making a biological condition an explicit input. The announcement describes that condition as a tissue, cell type, or cell state, selected from a reference catalog or generated from a researcher's own single-cell experiment through a cell encoder [vendor_claim] [1].
Known, claimed, inferred
AlphaGenome 2 changes the conditioning structure of the task rather than merely increasing output volume. Instead of producing fixed parallel outputs for bulk tissues and cell lines, it takes a DNA sequence together with one biological condition and predicts the effects for that condition [vendor_claim] [1].
Training covered bulk and single-cell samples from humans and mice, with more than 107 million individual cell profiles across 68 tissues [vendor_claim] [1]. The model predicts gene expression, histone modifications, chromatin accessibility, and DNA methylation; methylation is the additional readout relative to the earlier model [vendor_claim] [1].
The company says the model adapts to newly profiled cell states without retraining, including from small cell samples [vendor_claim] [1]. It also supports cross-modality imputation: given one measurement such as single-cell gene activity, it can predict an unmeasured property such as chromatin accessibility [vendor_claim] [1].
On TenK10K, described as a single-cell map of 28 primary immune cell types, Google DeepMind reports that AlphaGenome 2 identified causal variants at 30% compared with 16%, and directional gene effects at 43% compared with 23%, against the original AlphaGenome fine-tuned on single-cell data and other specialized tools [vendor_claim] [1]. The announcement also states that, on genome regions outside the training data, the model explains significantly more of the variation in gene expression across cell types within a tissue, but provides no quantitative figures for that result [vendor_claim] [1].
The announcement reports that the model computationally reproduced a published finding that a distant regulatory variant drives inflammatory bowel disease by overactivating ETS2 in gut immune macrophages [vendor_claim] [1]. The retrieved text ends mid-sentence in this example, so the full description of the result is unavailable [1].
Limitations & uncertainties
The evidence comes from an official Google DeepMind announcement rather than an independent evaluation. The TenK10K percentages reflect the vendor's own comparison and have not been assessed by a third party [1].
The announcement does not disclose the underlying benchmark protocol, uncertainty estimates, statistical significance values, or full definitions of the 30%, 16%, 43%, and 23% figures [1]. It does not give separate scores for the original AlphaGenome and each specialized tool, nor does it clarify whether all comparators used the same evaluation protocol [1].
Several operational and technical details are missing. The announcement does not state model architecture size, training compute requirements, or the underlying sequence-model architecture in reproducible detail [1]. Licensing terms, pricing, API rate limits, and commercial-preview eligibility criteria are also not provided [1].
Generalization beyond the described cell types and datasets is not established, and the retrieved excerpt is incomplete in the disease-variant example [1].
Evidence & sources
The key statements of this analysis, labeled by type, with the sources they rely on. Trendoviz does not run product tests.
- Primary-source or vendor statementGoogle DeepMind announced AlphaGenome 2 as a conditional model that predicts how DNA sequences and variants affect specific human cell types [vendor_claim] [9].[1]
- Primary-source or vendor statementThe model was trained on bulk and single-cell samples from humans and mice, with more than 107 million individual cell profiles across 68 tissues [vendor_claim] [9].[1]
- Primary-source or vendor statementAlphaGenome 2 predicts gene expression, histone modifications, chromatin accessibility, and DNA methylation, with methylation added relative to the earlier model [vendor_claim] [9].[1]
- Primary-source or vendor statementGoogle DeepMind says the model can adapt to newly profiled cell states without retraining and can infer an unmeasured property from a single supplied measurement [vendor_claim] [9].[1]
- Primary-source or vendor statementOn the TenK10K dataset, Google DeepMind reports causal-variant identification at 30% compared with 16% and directional gene effects at 43% compared with 23%, against the original AlphaGenome fine-tuned on single-cell data and other specialized tools [vendor_claim] [9].[1]
- Primary-source or vendor statementThe announcement states that the model computationally reproduced a published finding that a distant regulatory variant drives inflammatory bowel disease by overactivating ETS2 in gut immune macrophages [vendor_claim] [9].[1]
- Primary-source or vendor statementAlphaGenome 2 is available for non-commercial research through the AlphaGenome API and for commercial research in private preview on Google Cloud [vendor_claim] [9].[1]
Sources
- 1AlphaGenome 2: Deciphering how cells read DNA to better understand disease●deepmind.googleretrieved Nov 18, 2026
Updates & corrections
- First published · published automatically after passing every check
Found an error? Report it.