FreshSynapse

AI trends & daily updates, in plain English

AI Research

DeepMind's AlphaGenome 2 Reads DNA One Cell Type at a Time — and Finds Nearly Twice as Many Gene-Changing Variants

By FreshSynapse Editorial · · 4 min read

A scientist in a white lab coat and black gloves looking through a microscope in a laboratory

Photo: Hippopx (public domain). Illustrative image.

What happened. Google DeepMind announced AlphaGenome 2 on Tuesday 11 November 2026, its next-generation DNA sequence model. The first AlphaGenome predicted how a DNA change affects gene activity averaged across a tissue. The new model answers a narrower, harder question: what does this DNA change do in this cell type, in this state? That matters because the same genetic variant can be harmless in one kind of cell and a driver of disease in another — and until now, models mostly could not tell the difference.

How it works. AlphaGenome 2 is what DeepMind calls a conditional model. Alongside a stretch of DNA sequence, it takes one target condition — a tissue, a cell type or a cell state — and uses that condition at every stage of the network, so it learns how the same stretch of DNA is read differently in different cells. It was trained on more than 100 million individual cell profiles across 68 tissues, from human and mouse samples, turning static cell atlases into what DeepMind calls a predictive engine for human genetics — the same idea behind world models that learn how reality behaves. The model can also adapt to a scientist’s own single-cell data.

The headline numbers — with the caveat attached. The main result comes from TenK10K, a large single-cell map of 28 primary immune cell types. There, DeepMind reports that AlphaGenome 2 identified 30% of causal variants, against 16% for the original AlphaGenome fine-tuned on single-cell data and other specialised tools, and called the direction of a variant’s effect on its gene 43% of the time, against 23%. That is nearly twice as many gene-changing variants found — but it is also a reminder of how far there is to go: even the new model misses 70% of causal variants on this test. And every figure here is DeepMind’s own, from a technical paper that has not been peer reviewed. Treat the ranking as a strong signal, not a settled fact.

What you can actually use. AlphaGenome 2 is available from today for non-commercial research through the AlphaGenome API, with interactive Colab tutorials, and in private preview for commercial research. DeepMind names the Wellcome Sanger Institute as an ongoing collaborator. What it is not: a diagnostic tool. The announcement states plainly that both AlphaGenome models are intended for research use only and “have not been validated or approved for clinical or diagnostic purposes”. The announcement does not mention open weights or a specific licence.

Why it matters. Most genetic risk is not a broken gene — it is a small spelling change in the vast stretches of DNA that control when genes switch on and off, and those switches work differently in every cell type. A model that can point at the specific cell where a variant matters turns a statistical association into a testable biological lead: which cell to study, which pathway to probe, which experiment to run next. That is the promise. It joins a run of single models crossing old boundaries — one model now simulates sand, cloth, fluids and folding proteins — but the honest version is narrower: a research tool, free for academics, whose headline results still need independent replication. The genomics community will now do what it always does with a DeepMind release — try to break it, and see what survives. The same discipline applies to AI claims in the lab as anywhere else: a new test that asks AI to read real lab evidence found even the best models score only 66.7% — benchmarks are a starting point, not a verdict.

Sources

  • Google DeepMind
  • AlphaGenome 2
  • genomics
  • DNA
  • AI research

Related articles