Can computational biology accurately map how genetic mutations alter individual cell behaviors without requiring constant laboratory retraining?
Researchers at Google DeepMind have introduced AlphaGenome 2, an AI model designed to unravel how changes in our DNA affect different cells. As Google stated in a blog post published Wednesday, the system provides a much finer-grained view than broader models like AlphaGenome, which DeepMind released last year. By converting static biological databases into a predictive system, the architecture targets the exact points where genetic variations alter cell activity.
Predictive Resolution Across Cell Profiles
While nearly every human cell carries identical DNA sequence information, different cell types activate distinct gene programs depending on their activity and environment. AlphaGenome 2 captures these distinctions by evaluating how variations impact individual cell types and states. To build this capability, DeepMind trained AlphaGenome 2 on more than 100 million individual cell profiles across 68 tissues.
DeepMind reports that AlphaGenome 2 identifies nearly twice as many variants that can change gene expression in cells than its previous state-of-the-art model. For Adam Gayoso, a research scientist at Google DeepMind involved in the project, the primary advance centers on tying human genetics to cell biology.
"When you develop drugs, you’re targeting specific proteins and specific cell types," Google DeepMind’s genomics lead Žiga Avsec told The Verge.
Avsec explained that this specific resolution becomes critical when working to treat a particular disease. Building such an architecture only recently became feasible because new laboratory technologies can capture individual cells and their functional states at high resolution. In experimental evaluations, researchers found that AlphaGenome 2 could make predictions about new types of cells and biological states without needing to be retrained, while also highlighting gaps in experimental records to direct lab work. What economic trade-offs accompany this increased resolution?
The Computational Overhead of Precision
Higher precision directly increases processing expenditure. Avsec stated that AlphaGenome 2 is not replacing AlphaGenome any time soon because its more detailed predictions come at a computational cost and are more expensive to produce than those of its predecessor. Because of those resource demands, using AlphaGenome 2 to generate the complete genome-wide Atlas that DeepMind issued in September—mapping every single DNA letter change in the human genome—remained impractical.
This operational barrier separates broad baseline reference data from targeted inference. DeepMind is currently working to improve its genetics models by generating more data, including for cell states that are not represented well today, and Avsec noted that future versions could potentially be trained on data across multiple species.
"AlphaGenome 2 is the successor to DeepMind’s genome-predicting AI model AlphaGenome, which was released last year.
Google is making AlphaGenome 2 available for non-commercial research through the AlphaGenome API starting from Wednesday, alongside a limited private preview for commercial research. The release follows earlier diagnostic efforts: on 2026-05-06, Google DeepMind unveiled MedGemma 1.5 4B, a multimodal model marking a major leap from simple medical record analysis to interpreting complex visual data like CT scans, MRIs, and histopathology slides (we covered this on 06.05). Meanwhile, competitors are advancing parallel biomedical research projects, aiming to shorten the loop between digital predictions and practical tests.
Deploying high-resolution biological models now requires balancing computational budgets against the practical need for laboratory validation. Teams scaling predictive pipelines must decide whether to absorb the higher inference costs of cell-specific models or rely on cheaper genome-wide baselines until specialized validation is required.
What’s behind it
Where previous biological models worked with static databases, Google DeepMind's AlphaGenome 2 maps mutation effects directly to specific cell types. Will commercial partners in the restricted pilot manage to translate AlphaGenome 2's costly cell-level inference into economically viable targeted therapies?