On November 11, 2026, Google DeepMind announced AlphaGenome 2, an AI model that predicts the effects of DNA variants by cell type and cell state. It combines a DNA sequence of roughly 1 million base pairs with information about the cell, and predicts how the same variant affects different cells. Even when a disease-associated region of DNA is identified, it is not easy to pinpoint which variant, which gene, and which cell type are actually involved. The new model is intended to help address that problem. In an evaluation using immune cells, with the precision of selected candidates held at 90%, recall of likely causal candidates rose from 16% with the previous model to 30%. The result suggests the model could help narrow down experimental targets more efficiently, but the need to validate its predictions remains.
Combining DNA sequence with cell type and state
Cells in the human body basically carry the same DNA, but which genes are active, and how strongly, differs from cell to cell. The same variant may affect gene expression in one immune cell type and have almost no effect in another. Understanding what a disease-associated region of DNA does therefore requires looking not only at the base sequence but at what happens in which cells.
The original AlphaGenome took a DNA sequence of about 1 million base pairs as input and predicted a predefined set of biological readouts together. AlphaGenome 2, by contrast, accepts the target tissue or cell type, its state, and even the species (human or mouse) as input alongside the DNA sequence. By building these conditions into the model from its early layers, it produces predictions tailored to each cell.
Where in the model the conditions are introduced also affects accuracy. According to the technical paper by Adam Gayoso and colleagues, adding conditions only on the output side brought limited improvement. The team also compared the model against the original model fine-tuned on single-cell data, and reported that AlphaGenome 2 predicts cell-level gene expression and chromatin accessibility more accurately.
To specify a new cell state, the model uses gene expression data measured by RNA-seq. This data is converted into a representation of the cell state and fed into the model, allowing it to make predictions for cell states not seen during training, without additional training. In other words, the system examines the effects of variants from DNA sequence, while RNA measurements define the state of the cell being predicted. According to Google DeepMind's announcement, training used human and mouse tissue and single-cell data, covering 68 tissues and more than 100 million cell profiles. DNA methylation, which is involved in regulating gene activity, was also added as a prediction target.
Recall of causal candidates rises from 16% to 30%
To judge AlphaGenome 2's performance, it is necessary to check the comparison model and the conditions of the evaluation data. The research team compared AlphaGenome 2 with "AG1-sc," the original model fine-tuned on single-cell data, across 28 immune cell types in the TenK10K dataset. The evaluation used 6,346 positive examples, statistically selected from combinations of variant, gene, and cell type, plus the same number of control examples.
Positive examples were combinations whose statistically estimated probability of being a causal candidate for gene expression exceeded 0.9. Controls were combinations for which the same value was below 0.01, matched one-to-one with positives to account for the effect of variant position. The evaluation used chromosomes held out from training, and the TenK10K data itself was not included in the model's training.
In the evaluation pooling the 28 immune cell types, the team fixed precision at 90%. Precision is the proportion of the candidates the model selected as causal that were actually positives. Recall, on the other hand, indicates how many of all the positives the model managed to find.
Under this condition, AlphaGenome 2's recall rose from 16% for the previous model to 30%. That is an improvement of 14 percentage points, or a factor of 1.875. The "roughly double" improvement that Google DeepMind emphasizes is based on this comparison. However, the positives used in the evaluation are causal candidates narrowed down by statistical methods, not variants whose causal role was proven experimentally. The figure therefore cannot be interpreted as diagnostic accuracy for identifying causal variants across all diseases.
The paper also uses a 90% threshold under a different condition. When the accuracy of the "direction of effect," meaning whether a variant increases or decreases gene expression, was fixed at 90%, recall rose from 23% to 43%. This is a different metric from the evaluation that holds candidate precision at 90%, and the two should not be confused. Both results mean that, with a certain standard of prediction reliability maintained, more candidates can be extracted for study.
Examples in inflammatory bowel disease and T cells
When choosing experimental targets from disease-associated DNA regions, variant position alone is rarely enough to determine the cause. Several nearby variants may be inherited together, and the gene a variant affects can lie far away. With AlphaGenome 2, comparing predictions before and after introducing a variant makes it possible to examine which variant is likely to affect which gene in which cell type.
In the paper, the team tested whether the model could reproduce existing findings for the 21q22 region associated with inflammatory bowel disease (IBD). Among seven candidates examined in earlier experimental studies, it predicted the effect of the variant "rs2836882." The model linked it to the gene "ETS2," about 290,000 base pairs away, and predicted that the effect was confined to monocytes and macrophages. This is consistent with the mechanism revealed in existing experimental studies. However, it was an evaluation of whether the model could computationally reproduce a known phenomenon, not a case where the AI discovered a new cause and it was proven experimentally.
In another example, the team examined whether a variant's effect differs depending on the activation state of the same T cells. Using data from a COVID-19 mRNA vaccine study that was not used in training, they fed the model cell states derived from gene expression measured by RNA-seq. The model distinguished activated from resting T cells and predicted that a variant associated with severe COVID-19 has a stronger effect on gene expression in the activated state.
ITmedia NEWS has also reported that the model can make predictions for new cell states from a few dozen cells' worth of data obtained by researchers, without retraining. If both cell type and activation state can be specified, researchers may be able to choose the cells and conditions for experiments more efficiently. Still, the model's predictions about the relationship between variants and gene expression need to be confirmed in actual cells.
Drug discovery applications require independent validation
Genome-wide association studies (GWAS) are a method for statistically searching for DNA regions associated with disease. But finding an associated region does not reveal which variant within it acts on which gene in which cell type. Identifying drug targets requires narrowing down these causal relationships further. The research team precomputed effects under 510 cell and tissue conditions for 6,822,130 common single-nucleotide variants and other variants. This set of variants differs from the roughly 9 billion single-base substitutions covered by AlphaGenome Atlas, announced in September.
In his own assessment, Peter Donnelly, CEO of the genetic analysis company Genomics, noted that drug discovery requires identifying not only the target gene but also the direction of effect, meaning whether gene expression goes up or down, and the affected cell type. The company plans to integrate AlphaGenome 2 into its drug discovery platform "Mystra." However, it says it will evaluate with its own GWAS data how much useful information the new model's predictions add beyond existing genetic evidence such as distance to the gene, quantitative trait loci (QTL), and rare variants.
Donnelly also pointed out that of the roughly 6.8 million variants evaluated in the paper, high-confidence predictions were obtained for about 2%. That tally includes many common variants, and both the evaluation targets and the denominator differ from the result of recovering 30% of the positives selected in TenK10K. Variants without high-confidence predictions should not be read as having no effect; it means the model could not predict them with sufficient confidence.
The technical paper also describes the model's limitations. For example, effects of variants acting through poorly accessible chromatin regions, very distant regulatory elements, or post-transcriptional processes may be underestimated. Training data are also limited for cells exposed to external stimuli, cells in disease states, and transient cell states. Because the representation of cell state is learned from correlations among observed states, whether the model can accurately predict responses to gene perturbations or changes in cell state remains a subject for future validation.
Google DeepMind offers the AlphaGenome API for non-commercial research and is accepting applications for a limited preview for commercial research. However, the model has not undergone validation or approval for clinical or diagnostic use, and the timing of general availability of a commercial version has not been disclosed. The key question in judging AlphaGenome 2's practical value is whether it can newly identify causal variants, target genes, and cell types that existing genetic evidence alone could not reveal. How much it can contribute to drug discovery research will be determined through evaluation on independent data and experimental validation.