Google DeepMind has released AlphaGenome 2, a new AI model for genomics research that predicts how DNA variants affect specific human cell types — and finds nearly twice as many relevant gene changes as its predecessor, according to coverage by The Decoder of the company's announcement.
The release addresses one of genomics' stubborn problems: every cell in the human body carries the same DNA blueprint, but a heart muscle cell reads that DNA differently than an immune cell. Predicting those cell-specific differences — and pinpointing which of the millions of possible DNA letter changes actually cause disease — has remained one of the field's biggest unsolved challenges. Google DeepMind's announcement, covered widely on Thursday, positions AlphaGenome 2 as a direct attack on that problem, and Sundar Pichai highlighted the framing that "every cell in our body shares the same DNA blueprint" in the launch coverage reported by Benzinga.
From Bulk Predictions to Cell-Type Precision
The original AlphaGenome took a DNA sequence a million letters long and predicted its properties across thousands of experiments spanning many tissues at once. According to Google DeepMind, that parallel approach spread the model's capacity too thin, making it hard to capture subtle differences between closely related cell types.
AlphaGenome 2 works differently. It is a conditional model: alongside the DNA sequence, it takes a specific biological state as an extra input and computes effects for that one cell type. The system was trained on more than 107 million individual cell profiles spanning 68 tissues in humans and mice.
The model goes beyond predicting how strongly a gene is expressed. It also calculates how DNA is physically packaged inside the nucleus — showing which stretches of the genome are accessible — and now predicts DNA methylation, the chemical tags that act as molecular switches to silence or activate genes.
Benchmark Results: A Large Jump in Causal Variant Detection
On the TenK10K benchmark dataset covering 28 primary immune cell types, AlphaGenome 2 identified about 30 percent of causal variants and correctly predicted whether a variant upregulates or downregulates a gene in 43 percent of cases. A fine-tuned version of the original AlphaGenome managed just 16 percent and 23 percent on the same tasks, per Google DeepMind's figures reported by The Decoder.
That improvement matters because genome-wide association studies (GWAS) can only roughly locate disease signals within broad stretches of DNA. Better causal-variant prediction tells researchers which of the many variants in an associated region are actually worth experimental follow-up — the difference between a lead and a dead end.
The model can also infer missing data from a single measurement. If researchers measure only a cell's gene activity, AlphaGenome 2 provides predictions for additional properties, such as how tightly specific DNA stretches are packaged. That helps labs decide which wet-lab experiments are worth running next, and a key architectural advantage is adaptability: researchers can feed in gene activity data from just a few dozen cells, and the model computes effects for previously unseen cell states — often beating specialized models trained directly on that data, according to DeepMind.
Disease Mechanisms Reproduced Computationally
Two demonstrated cases show what the model can do without touching a lab bench. Working purely computationally, AlphaGenome 2 reproduced an inflammatory bowel disease mechanism previously identified in a study: it isolated the exact DNA variant in a regulatory switch far from the gene that selectively overactivates ETS2 in gut macrophages, while other cell types and genes remained largely unaffected.
In a separate test applied to data from a COVID-19 mRNA vaccine study, the model reliably distinguished vaccine-activated T cells from resting ones and showed that a gene variant linked to severe COVID-19 cases disrupts cell activity mainly in the activated state — an effect that would have remained invisible in resting cells.
Availability and What Comes Next
AlphaGenome 2 is available now for noncommercial research through an API, while commercial users get access through a closed preview. Google DeepMind is explicit that the system is not approved for clinical diagnostics — it is a research tool, not a medical device.
DeepMind has already precomputed predictions for hundreds of cell and tissue conditions, partly to support AI-driven drug development, and the company says it plans to roll out dedicated AlphaGenome 2 features on its agentic development platform Google Antigravity, where the model is set to be combined with tools like Gemini 4 for automated research workflows.
The release lands in a fast-heating AI-for-biology field. DeepMind's own AlphaFold lineage reset structural biology, and rivals are racing to apply large models to drug discovery. What distinguishes AlphaGenome 2 is the resolution: by conditioning on cell type, it moves variant interpretation from population-level averages to the specific cellular context where disease actually operates. If the benchmark gains hold up in independent use, the model could shorten the path from genetic association to mechanistic understanding — the most expensive step in modern drug development.
Stay Ahead of AIAI is reshaping drug discovery, genomics, and biotech faster than most coverage can keep up. For breaking AI news, research analysis, and the latest AI developments, AI Buzz Wire follows every breakthrough that matters.
Read more AI news →
