Google's New AI Can Find Where Disease Begins — in Your DNA

By Shakeel Ahmed — AI Growth Lab | November 12, 2026

Originally published on AI Growth Lab

On November 11, 2026, Google DeepMind released AlphaGenome 2, an AI model that predicts how a change in DNA affects gene activity in a specific cell type — a gut immune cell, a heart muscle cell, an activated T cell — instead of averaging across a whole tissue. Trained on more than 107 million individual cell profiles across 68 tissues, it identified nearly twice as many likely disease-causing variants as the original model. It is a research tool only: DeepMind says it is not validated or approved for clinical or diagnostic use.

That distinction — which cell, not just which gene — is the whole story. Most of the DNA changes linked to disease do not sit inside genes at all. They sit in the switches around genes, and a switch that matters in one cell type can be silent in another. Until now, AI could tell you a variant looked suspicious. AlphaGenome 2 tries to tell you exactly where it acts.

The problem: same DNA, different cells

Every cell in your body carries the same DNA. What makes a gut cell different from a brain cell is not the code — it is which parts of the code each cell reads. Predicting how a small DNA change ripples through that reading has been, in DeepMind's own words, an unsolved challenge and a key driver of human traits and disease.

The first AlphaGenome, previewed in June 2025 and published in Nature in January 2026, predicted how DNA variants change molecular activity across broad tissues. That was useful, but tissues are averages. A variant that drives disease in one rare immune cell can disappear entirely in a tissue average — and with it, the clue to a drug target.

AlphaGenome 2 closes that gap. DeepMind describes it as a step change:

"By zooming in from broad tissue averages to individual cell types and cell states, and redesigning the model's architecture, AlphaGenome 2 represents a step-change improvement in variant predictive performance." — Google DeepMind, "AlphaGenome 2: Deciphering how cells read DNA to better understand disease" (https://deepmind.google/blog/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease/)

What AlphaGenome 2 actually does

The model takes up to one million letters of DNA plus one chosen condition — a tissue, a cell type, or a cell state — and predicts how that sequence behaves in that condition. Its readouts include gene expression, chromatin accessibility, histone marks, and a new one: DNA methylation, a chemical tag linked to genes being switched off.

Three things make it different from its predecessor:

• It is conditional, not parallel. The original model predicted thousands of experiments at once. AlphaGenome 2 predicts for one condition at a time, feeding that condition into every stage of the network — so the same stretch of DNA is read differently in different cells, the way real biology works.
• It learns from your data, without retraining. A scientist can feed in their own single-cell experiment — as few as a few dozen cells — and the model's cell encoder adapts to it on the fly. Professor Sarah Teichmann of the University of Cambridge, co-founder of the Human Cell Atlas, noted in DeepMind's announcement that the model adapts directly to new single-cell data without retraining.
• It was trained at a scale genetics has never seen. More than 107 million individual cell profiles from human and mouse samples, covering 68 tissues, in bulk and single-cell form.

The numbers: nearly twice as many causal variants

DeepMind's headline test used TenK10K, a large single-cell map of 28 primary immune cell types. The results, reported by DeepMind:

• 30% of causal variants identified, versus 16% for the original AlphaGenome fine-tuned on single-cell data — nearly double.
• 43% directional accuracy — getting the direction of a variant's effect on its gene right — versus 23% before.
• More than 6.8 million variants pre-scored, with published scores for trait-linked variants across hundreds of human cell and tissue conditions, covering traits from height to type 2 diabetes.

Those figures are DeepMind's own, from its own benchmarks — no independent replication had appeared at the time of writing. Treat them as the company's claim, not a settled fact. But the direction is clear: cell-level resolution finds variants that tissue-level models miss.

This is the same lab that keeps pushing AI deeper into science. Just last week, DeepMind's AI found cracks in RSA encryption — we covered that here (https://dev.to/shakeelahmed/deepminds-ai-found-cracks-in-rsa-encryption-should-you-worry-4ooa) — and days ago, 1,000 AI agents built a working operating system in eight days — covered here (https://aigrowthlabshakeel.blogspot.com/2026/11/claude-agents-built-operating-system-8-days.html). AlphaGenome 2 is the biology version of the same trend: AI moving from answering questions about science to doing the science.

The proof it points to: a bowel disease gene, found in one cell

The most convincing demo is not a benchmark. It is a real disease, already solved by humans, that the model reproduced on its own.

A 2024 study in Nature traced inflammatory bowel disease to a variant in a distant regulatory switch that overactivates the ETS2 gene in gut macrophages, a kind of immune cell in the gut. AlphaGenome 2 picked out the same causal variant and showed that it selectively boosts ETS2 in those macrophages, with little change in other cell types or genes.

That is exactly the kind of result that matters for drug discovery. If you know the variant acts only in gut macrophages, you know which cell to target — and which cells to leave alone.

DeepMind reports a second check with the same pattern. Given data from a COVID-19 mRNA vaccine study, the model told vaccine-activated T cells apart from resting ones, and showed that a variant linked to severe COVID-19 disrupts gene activity more strongly in the activated cells — an effect that would be invisible in resting cells alone.

What it cannot do (yet)

Honesty matters more than hype with a health story, so here are the limits, stated plainly:

• It is not a medical tool. DeepMind is explicit: the AlphaGenome models are intended for research use only and "have not been validated or approved for clinical or diagnostic purposes." It will not read your personal genome, and it cannot diagnose anything.
• It still misses most variants. Even at 30%, the model recovers only a minority of the variants that statistics flag as likely causal. The paper is honest about that cap.
• It learns correlations, not proven cause and effect. A prediction is a lead for a lab experiment, not a verdict.
• It is expensive to run. Predictions cost considerably more compute than the original model, which makes a genome-wide atlas with this version impractical, according to DeepMind's genomics lead Žiga Avsec. The original AlphaGenome stays in service for the broad work.

Access reflects that research-only status. Non-commercial researchers can use AlphaGenome 2 today through the AlphaGenome API, with interactive Colab tutorials. Commercial research is in private preview. Genomics PLC, the Oxford-founded company, is already folding the model into its Mystra platform, and DeepMind says AlphaGenome 2 skills are coming to Google Antigravity, its agentic development platform.

What this means for medicine

AlphaGenome 2 will not cure anyone this year. What it does is shorten the hardest step in genetics: going from "this region of DNA is linked to a disease" to "this variant, in this cell, does this thing." That step is where most genetic discoveries stall, and where most drug programs die.

The work was built with a notably British cast of partners — Professor Sarah Teichmann at Cambridge, the Wellcome Sanger Institute, and Genomics PLC — on top of the Human Cell Atlas, the international effort to map every human cell type. That atlas gave the model its training data; the model turns the atlas from a map into a predictive engine.

If you want the big picture: AI in biology has moved from reading DNA to reasoning about it, cell by cell. Anthropic just split its best model in two — one for everyone, one you need permission to use — we covered that split here (https://aigrowthlabshakeel.blogspot.com/2026/11/claude-fable-mythos-split-access.html). DeepMind's move is the mirror image: take the most powerful model and point it at the smallest unit of the human body. The disease does not start in the organ. It starts in the cell. Now there is an AI that can look there first.

Sources

• AlphaGenome 2: Deciphering how cells read DNA to better understand disease — Google DeepMind, November 11, 2026 (https://deepmind.google/blog/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease/)
• Google DeepMind releases AlphaGenome 2 to predict how DNA changes affect specific cell types and disease — MLQ, November 11, 2026 (https://mlq.ai/news/google-deepmind-releases-alphagenome-2-to-predict-how-dna-changes-affect-specific-cell-types-and-disease/)
• AlphaGenome 2: DeepMind's DNA Model Now Works Cell by Cell — ShiftDelete, November 11, 2026 (https://shiftdelete.net/en/alphagenome-2)
• AlphaGenome 2: Google DeepMind Advances DNA Research — Vincy Powa News, November 11, 2026 (https://vincypowanews.com/google-deepmind-alphagenome-2-dna-cell-research/)