In November 2026, Google DeepMind announced AlphaGenome 2, an artificial intelligence model that represents a qualitative leap in the ability to predict how DNA variations affect specific cell types. The launch marks the direct evolution of AlphaGenome 1 — released in 2025 — and the AlphaGenome Atlas, which mapped all 9 billion possible single-letter changes in the human genome, but without specifying which cell type each variant would act upon.
The central problem of modern genetics is no longer simply identifying disease-associated variants; it is understanding why the same mutation has radically different effects depending on the cellular context. The same stretch of DNA is read differently by a heart muscle cell, a T lymphocyte, or a pancreatic beta cell. This cell-type specificity is exactly what AlphaGenome 2 was designed to decipher, and the result, according to its authors, is a step-change in predictive accuracy.
The team led by Žiga Avsec trained the model on more than 100 million individual cells across 68 different tissues. While AlphaGenome 1 received data collected in 'bulk' samples — that is, heterogeneous mixtures of many cell types — AlphaGenome 2 operates at single-cell resolution. This enables the model to learn how the same DNA sequence is interpreted differently by different cell types.
The new model also identified nearly twice as many variants that alter gene expression in cells compared to the previous state of the art. But the most notable feature may be its conditional architecture: instead of predicting properties in parallel for all tissues at once, AlphaGenome 2 is designed to predict the DNA effects for ONE specific biological condition — a tissue, cell type, or state — as input alongside the DNA sequence at every stage of its architecture.
This enables what the researchers call 'cross-modality imputation': if a scientist provides just one measurement, such as single-cell gene activity data, the model can predict unmeasured properties like chromatin accessibility. In other words, AlphaGenome 2 can fill experimental gaps that previously required weeks of laboratory work.
The emblematic case is inflammatory bowel disease (IBD). In 2024, researchers at the Francis Crick Institute identified a new risk factor: a DNA variant that makes the ETS2 gene hyperactive in immune cells. The problem was that this variant was located in a 'gene desert' region, far from protein-coding sequences, making its function virtually inexplicable by traditional methods. What did DeepMind do with AlphaGenome 2? It replicated aspects of this discovery in 30 minutes, without using any molecular biology techniques. "They managed to figure out the answer in 30 minutes, and it took us a year and a half of experiments," said James Lee, the immunologist who led the original study.
Available for non-commercial academic research via API and in private preview for commercial research, AlphaGenome 2 adds to DeepMind's growing family of genomic AI models that are transforming how science reads the human genome. The team is now working on training future versions on multi-species data and on feeding projects for 'virtual cells' — digital simulations that mimic entire biological systems.
Sources: The Verge, Nature, Google DeepMind
✓ Independent sources cross-checked and verified before publishing
