Google DeepMind Launches AlphaGenome 2 for Cell-Level DNA Prediction
Google DeepMind launches AlphaGenome 2, a genomic AI system designed to predict how the same stretch of human DNA behaves in particular cell types and biological states. The company says the model was trained on more than 107 million individual cell profiles spanning 68 tissues.
The release advances beyond the original AlphaGenome by making cellular context part of every prediction. Researchers can now ask how a variant may affect a gut macrophage, an activated T cell or another specified state instead of receiving only a broad tissue-level estimate.
The AlphaGenome 2 release adds four capabilities:
- Cell-type and cell-state conditioning across the model
- DNA methylation among the predicted molecular signals
- Generalization to cellular states not seen during training
- Research access through an updated AlphaGenome API
Google DeepMind Launches AlphaGenome 2 With Cell-State Conditioning
Human cells share nearly the same genome, but they activate different genes and regulatory programs. A variant that matters in one immune-cell state may have little effect elsewhere. AlphaGenome 2 combines a one-million-letter DNA sequence with a description of the target cell or state to model that difference.
According to Google DeepMind's announcement, the system retains the original model's sequence-level outputs and adds DNA methylation. Methylation is a chemical marker that helps regulate gene activity, making it a useful signal when researchers study how cells interpret the same underlying sequence differently.
The training data came from single-cell measurements, which capture molecular activity one cell at a time. That scale lets the model learn relationships between sequence and cell identity across healthy and disease-associated states, though its outputs remain computational predictions rather than laboratory findings.
How AlphaGenome 2 Changes the Genomics Workflow
A key feature is the ability to condition a prediction with a small snapshot of gene activity from a new cell state. DeepMind says researchers can supply single-cell RNA sequencing data from dozens of cells, allowing the model to extrapolate to that state without retraining the full system.
The AlphaGenome 2 preprint describes the system as a multimodal model that jointly predicts gene expression, chromatin accessibility, protein binding and other regulatory readouts. The authors report that this conditioning supports epigenomic imputation and variant-effect prediction in states absent from its training set.
DeepMind demonstrated the approach with vaccine-activated T cells. The model was given a small sample of cells measured after vaccination and then used to estimate the effect of a genetic variant associated with severe COVID-19. Its predicted effect was stronger in the activated state than in resting cells.
This is distinct from DeepMind's earlier AlphaGenome Atlas. That resource precomputed predictions for billions of variants using the first AlphaGenome model. AlphaGenome 2 changes the underlying inference process by letting a researcher specify the cellular context for a new sequence or variant.
The TenK10K and ETS2 Results Need Experimental Follow-Up
On the TenK10K single-cell benchmark, DeepMind reports that AlphaGenome 2 recovered 30% of causal variants, compared with 16% for the original model and specialized tools. It also predicted the correct direction of gene-expression effects for 43% of variants, versus 23% for those comparison systems.
The preprint presents the gain another way: recall of causal single-cell expression quantitative trait loci doubled at 90% precision. These are company-reported research results from a preprint. Independent replication and task-specific testing will determine how reliably the gains transfer to other datasets and disease areas.
One case study focused on inflammatory bowel disease and the ETS2 gene. Nature reported that the model reproduced a regulatory connection previously established by researchers led by James Lee, helping identify the gut macrophages in which the risk signal operates.
That example illustrates the intended role: narrowing a large search space before scientists run costly experiments. A credible prediction can prioritize a variant, cell type and mechanism, but it cannot establish that the mechanism causes disease or that targeting it will produce a safe treatment.
Research Access Comes With Clinical Limits
AlphaGenome 2 is available for non-commercial research through the AlphaGenome API. DeepMind is also accepting applications for a private preview aimed at commercial research. The company has not released the system as a clinical or diagnostic product.
The model's value will depend on how well its predictions hold up across ancestries, rare cell populations, experimental platforms and disease conditions. Single-cell datasets are expanding quickly, but they still contain measurement noise and uneven representation that can shape the patterns a model learns.
For drug developers and academic labs, the immediate opportunity is faster hypothesis generation. AlphaGenome 2 can suggest where a disease-linked variant acts and which molecular readouts to measure. Laboratory validation, clinical evidence and regulatory review remain necessary before any prediction can influence patient care.
Illustrative archival photograph: DNA sequencing researchers photographed by Linda Bartlett for the National Cancer Institute, via Wikimedia Commons, public domain. The image was cropped to 16:9 and resized; it does not depict AlphaGenome 2 or DeepMind's team.
