🔥 DeepMind Taught an AI to Read DNA Cell by Cell. Its Hit Rate Nearly Doubled.
Google DeepMind's AlphaGenome 2, out 11 Nov 2026, nearly doubled causal-variant hit rate (16% to 30%) by reading DNA one cell type at a time.

Your genome is identical in almost every cell you own. The thing that makes a neuron a neuron and a gut cell a gut cell is not the DNA itself but which stretches of it get read, in which cell, at which moment. A mutation that sits silent in your blood can flip a gene on in your intestine and quietly start a disease. Models that treat DNA as one flat ribbon of text walk straight past that.
On 11 November 2026, Google DeepMind released AlphaGenome 2, a DNA model built to stop walking past it (DeepMind). Here is the number that matters: on a benchmark called TenK10K, it correctly flagged 30% of the causal variants behind gene-expression changes, up from 16% for the previous version (The Neuron).
On the harder task of getting the direction of a variant’s effect right, it climbed from 23% to 43%.
Nearly double the hit rate, from a model you aim at one specific cell type instead of at the whole body. That targeting is the trick, and it is the entire story.
🧠 Why This Matters
For two decades, genome-wide studies have handed scientists a mountain of variants statistically linked to disease. The problem is that more than 90% of them sit in non-coding DNA, the regulatory switches between genes, and almost none come with an instruction manual. You know a letter is associated with, say, Crohn’s disease. You don’t know which gene it tweaks, in which tissue, or whether it turns that gene up or down.
That gap is where drug programmes stall. You cannot design a therapy around a switch when you cannot say what the switch does.
AlphaGenome 2 reads a one-million-letter stretch of DNA at once, the same span as the original (DeepMind). The difference is that it no longer answers “what does this DNA do” in general. It answers “what does this DNA do in this cell type,” and that specificity is what moved the numbers.
📊 Deep Dive
The original AlphaGenome produced a fixed set of parallel predictions: here’s the readout for this bulk tissue, here’s the one for that cell line, take it or leave it. AlphaGenome 2 is a conditional model. You feed it a DNA sequence and a target, a tissue, a cell type, a cell state, and it threads that condition through every layer of its architecture (DeepMind).
How the two models stack up:
- Training scale: the original leaned on bulk data; AlphaGenome 2 trained on over 107 million individual cell profiles across 68 tissues, from human and mouse samples (aiweekly).
- Causal-variant recall (TenK10K): 16% → 30% (DeepMind).
- Getting the direction right: 23% → 43%.
- New data: a built-in cell encoder lets it analyse fresh single-cell datasets without retraining.
- New readouts: it adds DNA methylation, and does cross-modality imputation, so one measurement (gene activity) can predict another (how open the chromatin is).
That last point is the quietly useful one. Wet-lab experiments are slow and expensive; a model that infers one measurement from another is a model that saves researchers a stack of them.
“What makes AlphaGenome 2 exciting is how it adapts directly to new single-cell data without retraining.” — Professor Sarah Teichmann, University of Cambridge (DeepMind)
As a sanity check, DeepMind pointed the model at a known result: the selective effect of a variant on the ETS2 gene in gut macrophages, a finding from a 2024 Nature study. The model reproduced it computationally (The Neuron).
⚠️ The Catch
Reproducing a finding you already knew is a useful check, not a discovery. It tells you the model isn’t hallucinating; it doesn’t tell you it will find something new.
And read that headline number the other way round: 30% found means 70% missed. Even DeepMind concedes there is “considerable room left to improve” on its own benchmark (The Neuron). The figures are also company-reported, on a single evaluation, with no independent replication yet.
Most important, this is not a diagnostic. DeepMind is blunt that AlphaGenome 2 is “intended for research use only and [has] not been validated or approved for clinical or diagnostic purposes” (DeepMind). Nobody is getting a genetic verdict out of this. Researchers are getting a better shortlist.
🎯 What Happens Next
The model is live now through the AlphaGenome API for non-commercial research, with commercial access in private preview (The Neuron). The immediate job it does is triage: of a thousand candidate variants, which handful deserve the expensive lab time, which gene to measure, which cells to use.
Expect the first real tests inside pharma pipelines, where cutting a shortlist from hundreds to tens is worth a lot of money. The genomics world will also want to see it run on variants nobody has characterised yet, not just the ones it can check its homework against.
🧩 Bigger Picture
This is the same playbook DeepMind ran with AlphaFold on protein structure: take a biology problem drowning in data, build a model that turns the data into predictions, hand it to researchers, and let the field move faster. The twist here is conditioning, the move from one-size-fits-all outputs to answers tuned to a specific cellular context.
Peter Donnelly, CEO of Genomics Limited, called it “an impressive advance in the field” (DeepMind). The honest framing is narrower and more interesting: the machines are getting good at telling your cell types apart, and your DNA finally gets read in context instead of in the abstract.
Biology has always known that location is everything. The software is just catching up.
Sources
- Google DeepMind — AlphaGenome 2: deciphering how cells read DNA to better understand disease (11 Nov 2026)
- The Neuron — Google’s AlphaGenome 2 Wants to Find Where Disease Begins
- AI Weekly — Google DeepMind Launches AlphaGenome 2, Trained on 107M Cell Profiles Across 68 Tissues
❓ Quick answers
What is AlphaGenome 2?
It is a DNA sequence model from Google DeepMind, released 11 November 2026, that predicts how genetic variants affect specific human cell types from a one-million-letter DNA input (DeepMind).
How much better is AlphaGenome 2 than the original?
On the TenK10K benchmark it identified 30% of causal variants versus 16% for the fine-tuned original, and got effect direction right 43% of the time versus 23% (DeepMind, 11 Nov 2026).
Can AlphaGenome 2 diagnose disease?
No. DeepMind states it is intended for research use only and has not been validated or approved for clinical or diagnostic purposes; it is available via the AlphaGenome API for non-commercial research.





