DeepMind's AlphaGenome 2 reads DNA cell by cell — and finds nearly twice the variants that switch genes on and off
Google DeepMind's AlphaGenome 2, announced November 11, predicts how genetic variants affect gene regulation in individual cell types rather than tissue averages — recovering 30% of causal variants at 90% precision on a single-cell immune dataset, nearly double its predecessor — and is available now for non-commercial research, though DeepMind stresses it is research-only and not validated for clinical use.
📊 Scorecard: our Bullish call on $GOOGL is scored from $293.45, the last close before it ran. Its first result comes after the next close.
The News
Google DeepMind announced AlphaGenome 2 on November 11, a next-generation model that reads DNA the way cells actually do — one cell type at a time. The original AlphaGenome predicted how genetic variants — small changes in the DNA code — affect gene regulation, the process that switches genes on and off. But it worked with tissue averages, blending together the many cell types in a sample. AlphaGenome 2 predicts those effects in individual cell types and states, which is how disease-linked variants usually act: a variant that matters in an immune cell may do nothing in a skin cell.
The model was trained on more than 107 million individual cell profiles across 68 tissues, a scale of single-cell data that lets it learn how the same stretch of DNA behaves differently depending on the cell reading it. On the TenK10K single-cell immune dataset — a collection of immune-cell data used to test these models — AlphaGenome 2 recovered 30% of putatively causal variants linked to gene expression at 90% precision, versus 16% for the original AlphaGenome fine-tuned on single-cell data. It also got the direction of the effect right — whether a variant turns a gene up or down — for 43% of variants, versus 23% for its predecessor. Put simply, it identifies nearly twice as many variants that change gene expression.
AlphaGenome 2 is available now for non-commercial research through the AlphaGenome API, with commercial research in private preview. DeepMind is clear about the limits: the model is for research only, and it is not validated or approved for clinical or diagnostic use. Even at 90% precision it recovers only a minority of the variants nominated as causal — 30% — so most of the genetic story behind a disease is still out of reach. All of the benchmark figures are DeepMind's own, reported from its own testing.
AlphaGenome 2 doubles a small number — and in genetics, doubling a small number is still progress.
Why It Matters
This matters because most of the genetic variants linked to disease sit outside the parts of DNA that code for proteins, in regions that control when and how strongly genes are switched on. Those control regions behave differently in different cells, so a model that averages across a whole tissue can miss exactly the variant that causes the problem — or flag the wrong one. A model that works cell by cell gets closer to how biology actually works, and that is what drug researchers need when they are hunting for the variant that drives a disease.
The jump from 16% to 30% recovery, and from 23% to 43% on effect direction, is a real improvement — but the numbers also show how far the field still has to go. Recovering 30% of causal variants at high precision means seven in ten are still missed. That is not a failure of the model so much as a measure of the problem: linking a DNA variant to a disease is hard, and no current model does it well. The honest read is that AlphaGenome 2 doubles a small number.
For Google, the release also extends its lead in AI for science — the same strand of work that produced AlphaFold, the protein-folding model that won a Nobel Prize for DeepMind's Demis Hassabis. AlphaGenome 2 is not a product with a price tag; it is a research tool, free for non-commercial use. Its value to Alphabet's shareholders is indirect: it burnishes DeepMind's scientific reputation, draws researchers into Google's ecosystem, and builds the case that Google's AI lead extends beyond chatbots into the lab.
Analyst's Take
Our call is Bullish. AlphaGenome 2 is a genuine step forward on a hard scientific problem, with nearly double the variant recovery of its predecessor on a standard test, and it is in researchers' hands now rather than promised for later. The limits are stated plainly by DeepMind itself — research-only, no clinical use, and still a minority of causal variants recovered — which makes the announcement easier to trust, not harder.
What to watch: whether independent labs reproduce the TenK10K gains on their own data, and how quickly the commercial private preview turns into paying research customers. A model that only its maker can validate is a claim; a model that other labs build on is a platform. The next evidence will come from papers that cite AlphaGenome 2, not from DeepMind's blog.
Are you bullish or bearish on this story?
By The Numbers
FAQ
What is a genetic variant?
A genetic variant is a small change in a person's DNA code — a single letter swapped, added or deleted. Most are harmless, but some change how genes work, and those changes can raise the risk of disease.
What does AlphaGenome 2 do?
It predicts how a genetic variant affects gene regulation — whether it switches a gene on or off, and by how much — in a specific type of cell, rather than averaging across a whole tissue. It is a research tool for scientists studying the genetic roots of disease.
How much better is it than the original?
On the TenK10K single-cell immune dataset, DeepMind reports it recovered 30% of putatively causal variants at 90% precision, versus 16% for the original AlphaGenome, and got the direction of the effect right for 43% of variants versus 23% — nearly twice as many in both cases.
Can doctors use it to diagnose patients?
No. DeepMind stresses AlphaGenome 2 is research-only and is not validated or approved for clinical or diagnostic use. It is a tool for scientists, not a medical test.
What is 'precision' here?
Precision is the share of the variants the model flags that turn out to be genuinely linked to a change in gene expression. At 90% precision, about nine in ten flagged variants are real — but the model still finds only 30% of all the causal variants out there.
👁
Source: Google DeepMind