Skip to content
trendoviz

AlphaGenome 2 predicts variant effects by cell type

Google DeepMind has announced AlphaGenome 2, a conditional DNA sequence model designed to predict how variants affect specific human cell types. Its reported performance gains come from the company's own evaluation rather than an independent benchmark.

29 out of 100Emerging7D: Insufficient history ProvisionalPublished

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

  1. 1

Updates & corrections

  1. First published · published automatically after passing every check

Found an error? Report it.