Google DeepMind unveiled AlphaGenome 2 on Wednesday, an artificial intelligence model that predicts how DNA sequence variations alter gene activity inside specific human cell types and disease states. The system analyses one biological condition at a time, allowing researchers to trace how genetic mutations cause disease in specialised cells that broader tissue models often overlook.
The model is available immediately for non-commercial research through an application programming interface, with commercial access offered in a private preview. It arrives as AI laboratories expand their biological research tools, shifting from cataloguing static DNA sequences to predicting how genetic code functions across living tissues.
A conditional architecture for individual cells
The original AlphaGenome model, released last year, predicted molecular activity across broad tissues and cell lines in parallel from a one-million-letter DNA sequence. While fast, that approach spread the network across thousands of simultaneous measurements, obscuring differences between closely related cell types where subtle mutations exert their effect.
To resolve those differences, DeepMind rebuilt the network as a conditional model, according to an announcement published by Google DeepMind. Instead of evaluating multiple tissues at once, AlphaGenome 2 takes a target biological condition, such as a specific cell type, tissue or cellular state, and feeds that identity through every layer of the architecture alongside the DNA sequence.
DeepMind trained the system on bulk and single-cell profiles from human and mouse tissue samples, incorporating more than 107 million individual cells across 68 tissues. Alongside gene expression, RNA splicing and chromatin accessibility, the model predicts CpG DNA methylation, a chemical modification that controls gene activity.
The architecture also carries out cross-modality imputation. If a scientist uploads a single measurement, such as gene expression from an RNA-sequencing experiment, the model can infer unmeasured properties such as chromatin accessibility, helping laboratories decide which physical tests to conduct next.
Pinpointing disease variants in immune cells
By evaluating specific cell identities, the model identifies variants that only become active under certain conditions. In benchmark tests on TenK10K, a dataset covering 28 primary immune cell types, AlphaGenome 2 identified 30 per cent of causal expression-linked variants at 90 per cent precision, compared to 16 per cent for its predecessor, Unite.AI reported. The model also reached 43 per cent accuracy in predicting the direction of gene effects, nearly double the 23 per cent scored by the previous model.
In the company’s technical paper, researchers detailed how AlphaGenome 2 traced a known regulatory switch for inflammatory bowel disease. The system examined a candidate set of seven variants, singled out the variant rs2836882, and showed that it boosted the ETS2 gene specifically in gut macrophages while leaving other immune cells unaffected.
The model can also evaluate cell states it has never seen during training without needing fine-tuning. When supplied with single-cell RNA snapshots from a COVID-19 vaccine trial, AlphaGenome 2 separated activated T cells from resting T cells. It predicted that a mutation associated with severe COVID-19 disrupted gene regulation far more heavily in activated immune cells than in dormant ones.
Drug discovery and biological modelling
Connecting population genetics to individual cell types addresses a recurring bottleneck in pharmaceutical development. Genome-wide association studies routinely link stretches of DNA to illnesses such as diabetes or heart disease, but identifying which variant within that stretch acts on a given cell type usually requires years of laboratory screening.
Žiga Avsec, genomics lead at Google DeepMind, told The Verge that cell-level resolution is essential for treatment design. “When you develop drugs, you’re targeting specific proteins and specific cell types,” Avsec said.
To support research teams without large computing clusters, DeepMind precomputed predictions for common genetic studies. The release includes a database mapping regulatory enhancers to 39,323 genes across 640 primary single-cell conditions, along with scores for 6.8 million common variants across 510 human conditions. Peter Donnelly, chief executive of Genomics PLC, said his company plans to integrate AlphaGenome 2 into its Mystra drug-discovery platform to turn genetic association signals into candidate drug targets.
In the longer term, researchers see these sequence models contributing to broader biological simulations, known as virtual cells, which aim to replicate internal cellular processes digitally.
Computational limits and experimental caveats
The precision of conditional modelling introduces practical constraints. Because AlphaGenome 2 evaluates each cell type and state separately, running it is computationally more demanding than generating broad tissue predictions. Avsec said these costs mean AlphaGenome 2 will not immediately replace the broader AlphaGenome Atlas, which mapped all nine billion single-letter changes in the human genome across generalised tissues.
DeepMind’s researchers noted several biological limits in their paper. AlphaGenome 2 still recovers only a minority of the causal variants suggested by statistical genetics. The model can underestimate mutations located outside accessible chromatin regions, variants in distant regulatory elements, or effects that alter biological processes after transcription.
The training data also relies on steady-state observational measurements rather than active cellular changes, capturing correlations rather than direct cause and effect. It remains unproven whether the model can accurately forecast how cells react to gene knockouts or physical transitions between states.
DeepMind plans to release AlphaGenome 2 Skills on its Science on Antigravity development platform, pairing the model with agentic software for experimental workflows. Both AlphaGenome models remain restricted to research use and have not been validated or cleared for clinical diagnosis.