Google DeepMind Launches AlphaGenome 2, a Conditional Model for Cell-Type-Specific Variant Prediction

Google DeepMind announced AlphaGenome 2 on November 11, 2026, a deep-learning model that predicts how DNA variants affect gene regulation in specific human cell types and states. The model is available today for non-commercial research via the AlphaGenome API, with commercial access in private preview [1].

From bulk tissues to single cell types

The original AlphaGenome, released in 2025, predicted molecular activity across bulk tissues and cell lines in parallel. That design "spread the model's capacity thin and made it hard to capture subtle differences between closely related cell types," DeepMind says [1]. AlphaGenome 2 instead takes a 1-million-base DNA sequence and a target biological condition—tissue, cell type, or cell state—as joint inputs, conditioning the prediction on that condition at every layer of its U-Net-style architecture [1][7].

The model was trained on over 107 million individual human single-cell profiles, plus mouse pseudobulk data, across 68 tissues, along with bulk functional-genomics data and cell-type-resolved DNA methylation [1][7]. Its readouts include gene expression, chromatin accessibility, histone modifications, transcription-factor binding, RNA splicing, and, new to version 2, CpG DNA methylation [1][7].

Benchmarks: DeepMind's own evaluations

DeepMind reports that on the TenK10K dataset—a single-cell map of 28 primary immune cell types—AlphaGenome 2 identified 30% of putatively causal expression-linked variants at 90% precision, versus 16% for AlphaGenome 1 fine-tuned on single-cell data. Direction-of-effect recall rose from 23% to 43% at 90% accuracy [1][4][6]. On held-out genome regions, the model explained a median 34% of within-tissue, cell-type-specific expression variance, up from 23% for the fine-tuned predecessor [4].

These figures come from DeepMind's own paper, for which the company has not named a journal or preprint server [1][6]. Independent computational biologist Peter Koo noted "the engineering that went into it and the performance is quite impressive," while adding that the underlying ideas are not novel [2].

Disease examples: IBD and COVID-19

DeepMind highlights two cases where AlphaGenome 2 reproduced or extended experimental findings.

In inflammatory bowel disease, a 2024 study led by James Lee at the Francis Crick Institute used gene editing to show that a variant in a distant enhancer overactivates the ETS2 gene in immune cells. AlphaGenome 2, given only the DNA sequence and cell-type condition, singled out the same causal variant from a seven-variant credible set and predicted the effect specifically in gut monocytes and macrophages [4]. Lee said the model "managed to figure out the answer in 30 minutes, and it took us a year and a half of experiments" [2].

Using single-cell data from a COVID-19 mRNA vaccine study that the model had not seen during training, AlphaGenome 2 distinguished vaccine-activated CD8+ T cells from resting ones. It predicted that a variant linked to severe COVID-19 disrupts gene activity more strongly in the activated state—an effect that would be diluted in unactivated cells [1][4].

Generalizing to unseen cell states

A cell encoder lets researchers upload raw single-cell RNA-seq data—even from a few dozen cells—to generate a condition embedding without retraining. DeepMind says this expression conditioning matches or outperforms models fine-tuned directly on the new dataset [1][7]. The company also precomputed two resources: enhancer–gene links across 640 primary cell conditions for 39,323 genes, and variant-effect scores for roughly 7 million common and trait-associated variants across 510 human conditions [1][4].

Limits and caveats

DeepMind emphasizes several constraints. AlphaGenome 2 recovers only a minority of the expression-linked and disease-associated variants that statistical analyses nominate as likely causal [4]. Sign prediction AUROC drops to 0.63 for variants outside accessible chromatin [7]. The model learns correlations from observational steady-state profiles, not causal relationships; it is unclear whether it can predict responses to knockouts, overexpression, or dynamic state transitions [4]. Neither AlphaGenome version has been validated or approved for clinical or diagnostic use [1].

The conditional architecture also costs more compute per prediction. DeepMind's Žiga Avsec told The Verge that this made a genome-wide atlas with AlphaGenome 2 impractical, which is why the AlphaGenome Atlas released in September used the original model [6]. Model weights are not yet released; the paper notes they will follow final publication [5].

Broader context

Genomics Ltd CEO Peter Donnelly said his company plans to incorporate AlphaGenome 2 into its Mystra drug-discovery platform, calling sequence-prediction models "an important tool in helping bridge that gap" from GWAS signals to drug-programme decisions [1]. DeepMind is collaborating with the Wellcome Sanger Institute to profile more cell types, developmental lineages, and perturbed states [1].

The field is moving toward conditioning on cell state [2]. AlphaGenome 2's scale—over 100 million training cells, enhancer–gene links across 640 primary cell conditions, and variant scores across 510 human conditions—and its zero-shot generalization to new single-cell data are distinctive [1][4][7]. Whether the reported gains translate into faster experimental prioritization remains to be seen by the wider community.