Google DeepMind’s AlphaGenome 2 sharpens DNA predictions by cell type

• DeepMind reports better predictions of DNA variant effects in specific cell types, including newly profiled states without dataset-specific fine-tuning.
• In an immune-cell benchmark, recall of putatively causal expression variants rose from 16% to 30% at 90% precision.
• Free noncommercial API access and precomputed predictions open the tool to researchers; clinical use remains unvalidated.

Google DeepMind, Google’s AI research unit, announced AlphaGenome 2 on November 11, extending its DNA prediction system to the specific cell types and states in which genetic changes act. The research tool could help scientists choose which disease-associated variants and target genes to investigate, but its benchmark gains do not establish that it can predict the outcome of laboratory interventions.

Company-reported gene-expression comparison across blood-cell differentiation: observations (left), AlphaGenome 2 (middle) and AlphaGenome 1 fine-tuned on single-cell data (right). Colors show Z-scored expression. Graphic: Google DeepMind. Source: Google DeepMind.

Why the cell matters

AlphaGenome 2 is a model for predicting gene regulation: how DNA sequences influence molecular activity, including which genes are switched on and how strongly. DeepMind, formed in 2023 by combining DeepMind and Google Research’s Brain team, launched the original AlphaGenome in 2025. It subsequently released AlphaGenome Atlas, containing predictions for all 9 billion possible single-letter changes in the human genome.

That breadth left an important resolution problem. A tissue sample mixes cells with different jobs, and a genetic variant can act in one population while having little effect in another. Non-coding DNA, which does not encode proteins but can contain switches controlling gene activity, makes up 98% of the human genome, as independent reporting on the launch explains.

James Lee, an immunologist and geneticist at the Francis Crick Institute who led a 2024 study identifying an inflammatory bowel disease mechanism, told that publication that the original model’s limited cell-type specificity had constrained its usefulness to his team. That is the practical problem the update seeks to address.

The new model takes a million-letter DNA sequence together with a specified tissue, cell type or cell state. Rather than generating a fixed collection of tissue outputs in parallel, it uses that biological condition throughout its architecture and predicts one condition at a time, DeepMind said in its announcement.

The sequence model was trained on human and mouse bulk and single-cell datasets. A separate cell-expression encoder was trained on a corpus of 107 million human single-cell gene-activity profiles spanning 68 tissues. A post-training stage aligned those representations with the frozen sequence model. Researchers can then supply their own single-cell RNA-sequencing measurements to describe a new cell population without fine-tuning the model for that dataset.

The system also predicts unmeasured properties, such as chromatin accessibility—how exposed DNA is to regulatory machinery—from available measurements. It adds DNA methylation to the molecular readouts supported by the earlier model. These capabilities could help select the next experiment when profiling every molecular feature of a scarce cell population is impractical.

Company-reported TenK10K results across 28 immune-cell types: recall at 90% precision and the fraction retained at 90% sign accuracy. These are separate tasks, not overall accuracy rates. Chart: Magica; data: Google DeepMind. Source: AlphaGenome 2 research paper.

What the near-doubling measures

DeepMind’s “nearly twice” claim concerns recall at a demanding confidence threshold, rather than a doubling of accuracy across all tasks. In the company’s research preprint, the TenK10K expression benchmark covered 6,346 fine-mapped variant–gene–cell-type effects across 28 primary blood immune-cell types. These are effects associated with changes in gene expression, designated putatively causal through statistical fine-mapping. The work was funded by Google DeepMind and conducted by its employees.

The first comparison tested fine-mapped effects against equal numbers of negative controls matched for regulatory-region overlap and distance to the target gene’s transcription start site. Precision measures how many selected predictions were correct within that constructed test; recall measures how many positive examples the model recovered. Thus the 30% result still leaves most benchmark positives unrecovered at that threshold. It should not be read as a 90% success rate for finding disease causes in an unfiltered genome.

For the direction task, the model ranked the fine-mapped effects by predicted magnitude and retained the highest-ranked fraction at the stated sign accuracy. Across the full precision–recall curve for the first task, the improvement was more modest: its area rose from 0.72 to 0.76. The near-doubling describes a particular operating point.

A separate evaluation on held-out genomic regions supports the cell-resolution claim: across human and mouse tissue comparison groups, AlphaGenome 2 explained a median 34% of within-tissue variation in cell-type-specific gene expression, versus 23% for the fine-tuned predecessor.

Existing tools remain relevant

The predecessor was a strengthened baseline, fine-tuned on single-cell data. The paper also compares AlphaGenome 2 with Scooby, a model of single-cell genomic profiles, and CREsted, a tool for modeling cell-type-specific enhancers. Cell-conditioned models already existed; DeepMind’s claim is better prediction and broader generalization, supported by its comparisons, rather than the invention of cell-specific sequence modeling.

The gains are uneven. On several bulk-tissue benchmarks for splicing, chromatin accessibility, DNase sensitivity and polyadenylation, the paper reports performance comparable to AlphaGenome 1. On an unseen brain dataset from 190 donors with and without Parkinson’s disease, however, the authors report better variant-effect predictions than cohort-trained CREsted and cohort-fine-tuned Scooby. The comparison used chromatin-accessibility effect direction for CREsted and gene-expression effect direction for Scooby. The latter relied on Scooby’s published predictions at its reporting threshold, rather than a full sweep of operating points.

Another comparison narrows any claim of general superiority. On TraitGym, a complex-trait benchmark with 11,400 variants and nine matched controls per positive example, the authors said AlphaGenome 2 matched GPN-Star, a predictor based on evolutionary conservation. Combining the two outperformed either alone, supporting the use of complementary signals when prioritizing variants.

Recovering a known disease mechanism

A useful demonstration concerns inflammatory bowel disease. Experiments published in 2024 had established that a regulatory variant increases activity of the ETS2 gene in immune cells, despite lying roughly 290,000 DNA letters away from that gene. AlphaGenome 2 retrospectively selected the implicated variant from seven candidates, linked it to ETS2 and localized the predicted effect to monocytes and macrophages, according to the paper.

That result shows why identifying the nearest gene or averaging across a tissue can be insufficient. It is a computational recovery of an experimentally established mechanism, rather than a newly proven treatment target. The authors state that they conducted no new wet-lab experiments for the study.

The model also used previously unseen data from a COVID-19 vaccination study to distinguish activated immune-cell states. Its predictions nominated CXCR6, a gene encoding a chemokine receptor, as a target of a variant associated with COVID-19 severity, with stronger effects in particular activated T-cell populations. That remains a predicted regulatory mechanism requiring experimental scrutiny.

Free access, with a computational tradeoff

AlphaGenome 2 is available through the AlphaGenome API for noncommercial research, with commercial research in private preview. The documentation says noncommercial API use is free, query rates vary with demand, and live predictions suit analyses involving thousands of predictions rather than millions. Atlas lookups permit higher rates.

The redesigned architecture has a cost: the paper says a separate prediction is required for each condition, making it more computationally intensive than fixed-output multitask models. Finer biological resolution therefore increases the work required to scan many cell states.

The researchers precomputed scores for about 6.8 million common and trait-associated variants across 510 human conditions, alongside an enhancer–gene linking atlas across 640 primary human conditions. The downloadable release is filtered by predicted effect sizes and enhancer-link thresholds. These selected variant scores serve a different purpose from the earlier Atlas’s exhaustive single-letter coverage; the headline totals describe different resources.

For now, access also depends on the release channel. The paper promises model source code and weights upon final publication. DeepMind says it will additionally bring AlphaGenome 2 Skills to its Antigravity development platform, extending its planned distribution beyond the current API and tutorials. Commercial pricing and a timetable for broader commercial access are not specified in the launch materials.

What experimental validation must resolve

The authors acknowledge that the model can underestimate effects outside accessible chromatin, at very distant regulatory elements and through mechanisms acting after transcription. Transient, stimulated and disease states remain underrepresented, and closely related neuronal populations can still be difficult to distinguish.

There is a deeper limitation: the cell representations come from observational profiles and encode correlations. The paper says it remains unclear whether AlphaGenome 2 can predict responses to gene knockout, overexpression or dynamic changes in cell state. Generalizing to a new observational dataset does not establish that it can predict what happens when a scientist intervenes.

DeepMind says it is working with collaborators to profile more cell types, developmental lineages and perturbed cell states. The consequential test for researchers is whether model-nominated variants and target genes hold up in prospective experiments, particularly where existing data are sparse. DeepMind says AlphaGenome 2 and its outputs have not been validated or approved for clinical or diagnostic use. As coverage of the launch notes, the company has announced no date for medical use.