Human genetics is the same book written in every cell of the body — heart, brain, immune system. The difference is which pages each one reads. This apparent simplicity conceals a decades-old problem: we know that most disease-linked genetic variants reside in the non-coding regions of DNA, but we cannot map with precision which cell or tissue each mutation actually affects. It's like knowing that a single word changed in the middle of a three-billion-letter novel, without knowing which paragraph that alteration affects.
Google DeepMind published AlphaGenome 2 today, an AI model that solves exactly this problem. Unlike AlphaGenome 1 (released in 2025) and the Atlas (September 2026), which produced predictions averaged across entire tissues, AlphaGenome 2 is a conditional model: you provide a DNA sequence and, in parallel, tell the model which cell type or biological state you want to analyze. The model then computes the effect of that specific variant in that specific cell.
The model was trained on more than 107 million individual cell profiles, distributed across 68 human and mouse tissues. The architecture was redesigned from the ground up — instead of producing multiple parallel outputs like AlphaGenome 1, AlphaGenome 2 focuses on a single condition at a time, which allows it to learn how the same stretch of DNA is read in radically different ways depending on cellular context.
The practical results are striking. On the TenK10K benchmark, which maps 28 primary immune cell types, AlphaGenome 2 identified approximately 30% of causal variants — nearly double what the previous model achieved. And it's not just quantity: when the model has high confidence, it also predicts whether the variant increases or decreases gene activity, information that is crucial for drug development.
James Lee, immunologist and geneticist at London's Francis Crick Institute, described a concrete case. His team spent a year and a half of gene-editing experiments to identify potential new treatments for inflammatory bowel disease. AlphaGenome 2 replicated aspects of that discovery in 30 minutes — without a single pipette being lifted. "It's reassuring from our perspective," Lee told Nature.
The model goes beyond what it was trained on: researchers can feed AlphaGenome 2 single-cell RNA-seq data from a new disease state — a tumor, a fibrotic tissue — and the model generates variant effect predictions for that state, without retraining. Adam Gayoso, DeepMind scientist and co-author of the study, explained that this accelerates the prioritization of variants underlying rare and complex diseases.
Beyond gene expression, histone modifications, and chromatin accessibility — the same metrics as AlphaGenome 1 — the model now also predicts DNA methylation, a feature widely requested by the scientific community. The previous Atlas contained predictions for all 9 billion possible single-letter changes in the human genome; now each of those 9 billion predictions can be refined to the correct cellular context.
AlphaGenome 2 is available via API for non-commercial research and in private preview for commercial research. DeepMind also announced that AlphaGenome 2 Skills will soon be released on the Antigravity platform, Google's agentic development tool, combining specialized models with AI agents to accelerate scientific discovery.
The next step, according to DeepMind, is to expand the spectrum of human and developmental cell types in collaboration with the Wellcome Sanger Institute. As single-cell mapping of the human body becomes routine, models like AlphaGenome 2 will become even more powerful — turning static atlases into predictive engines of human genetics.
What remains is the question of how far we can go with predictive AI before hitting the limits of real biology: the model can predict variant effects in new cellular states, but experimental validation remains the only path to confirm whether the prediction matches biological reality. The distance between computational prediction and clinical therapy is enormous, and AlphaGenome 2, as impressive as it is, remains a magnifying glass — not a treatment.
Sources: Google DeepMind Blog, Nature, The Verge
✓ Independent sources cross-checked and verified before publishing
