Every cell in your body carries the same DNA, but a heart muscle cell reads it very differently from an immune cell. That difference is why most disease-linked DNA variants have been so hard to pin down — until now, models predicted the average across a whole tissue. On November 11, 2026, Google DeepMind announced AlphaGenome 2, a model that predicts what a DNA variant does in one specific cell type at a time. I read DeepMind's official announcement line by line and cross-checked the numbers against independent same-day coverage. Here is what it does, what the numbers really mean, and the catch.
Image: Google DeepMind
How I checked this
I read DeepMind's November 11 announcement, "AlphaGenome 2: Deciphering how cells read DNA to better understand disease," in full, then compared its claims with independent coverage from MLQ and The Verge. All figures below are DeepMind's own, from its announcement and technical paper — no independent lab has replicated them yet.
What AlphaGenome 2 actually does
AlphaGenome 2 takes a stretch of DNA up to one million letters long, plus one chosen condition — a tissue, a cell type, or a cell state, like a T cell activated by a vaccine — and predicts how that DNA behaves in that condition: which genes switch on, how the DNA is packaged, and how it is marked. It also predicts DNA methylation, a readout the first version did not offer. The trick is a "cell encoder": the model was trained on more than 107 million individual cell profiles from human and mouse samples across 68 tissues, so it can adapt to a scientist's own single-cell experiment — with as few as a few dozen cells — without retraining. DeepMind has had a loud November: the same lab also put a humanoid robot on a live Mercedes-Benz factory line with Gemini Robotics 2.5.
The headline numbers — and what they don't prove
On TenK10K, a large single-cell map of 28 primary immune cell types, AlphaGenome 2 identified 30% of putatively causal expression variants at 90% precision, against 16% for the first AlphaGenome fine-tuned on single-cell data — nearly twice as many. It also got the direction of the gene effect right 43% of the time, versus 23%. Those are real gains on DeepMind's own benchmark. But flip the numbers: at 90% precision, seven in ten causal variants are still missed. And the model costs more compute per prediction than version 1 — DeepMind's genomics lead Ziga Avsec told The Verge that made a genome-wide atlas with AlphaGenome 2 impractical, so version 1 stays in service.
Who can use it — and who can't
AlphaGenome 2 is available today for non-commercial research through the AlphaGenome API, with interactive Colab tutorials; commercial research is in private preview, with no public price. And here is the line DeepMind repeats twice: both AlphaGenome models are for research use only and "have not been validated or approved for clinical or diagnostic purposes." This is not a tool that reads your personal genome — and anyone selling it as one is misrepresenting it. The compute behind models like this keeps growing, too: Google launched four TPU systems in a single year.
Why this matters — our take
Most disease-linked variants sit outside genes, in DNA regions that switch genes on and off — and those switches work differently in every cell type. A model that reads DNA cell by cell turns static cell atlases into a predictive engine: researchers can ask which cell type a variant acts in before running a single experiment. That is a genuine step toward finding the cellular drivers of disease. AI is moving into real biology fast — OpenAI's GPT-6 Astra took a real wet-lab exam the same week, scoring 66.7% on a 100-task benchmark. My verdict: this is the most useful kind of AI biology news — a real capability with published numbers and stated limits, not a demo. The limits matter as much as the gains: 30% is not 90%, the compute cost is real, and clinical use is explicitly off the table. Watch next: whether independent labs reproduce the TenK10K gains, and whether the promised AlphaGenome 2 Skills for Google's Antigravity platform make the model usable inside agent workflows. Google's AI even answered from space this month — the lab's ambition is clearly not slowing down.
AlphaGenome 2 will not diagnose anyone — but it may help scientists find which cell a disease variant actually acts in, and that is where cures start. What would you want AI to solve first in biology — rare diseases, cancer, or ageing? Tell us in the comments.
Sources
- Google DeepMind, "AlphaGenome 2: Deciphering how cells read DNA to better understand disease," November 11, 2026 — https://deepmind.google/blog/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease/
- MLQ.ai, "Google DeepMind releases AlphaGenome 2 to predict how DNA changes affect specific cell types and disease," November 11, 2026 — https://mlq.ai/news/google-deepmind-releases-alphagenome-2-to-predict-how-dna-changes-affect-specific-cell-types-and-disease/
- ShiftDelete, "AlphaGenome 2: DeepMind's DNA Model Now Works Cell by Cell," November 11, 2026 — https://shiftdelete.net/en/alphagenome-2


Comments
Post a Comment