AI/News

AlphaGenome 2: DeepMind’s AI Spots Disease Genes by Cell Type

DeepMind’s new AI model reads DNA one cell type at a time, and says that focus nearly doubles its hit rate for finding disease-linked genetic variants.

A gene-expression comparison chart from Google DeepMind showing AlphaGenome 2 predictions across different human cell types.
A chart comparing AlphaGenome and AlphaGenome 2 predictions across cell types. Image: Google DeepMind.

Google DeepMind has released AlphaGenome 2, a new AI system that predicts how a single strand of DNA behaves inside a specific cell type, and the lab says it identifies nearly twice as many disease-linked gene variants as its predecessor. The model, announced on the DeepMind blog on November 11, 2026, is available now through the AlphaGenome API for non-commercial research, with a private preview for commercial use opening through Google Cloud.

AlphaGenome 2 is a successor to AlphaGenome, the genome-reading model DeepMind first released in June 2025 and later published in Nature in January 2026. Where the original model predicted gene activity across many tissues at once, AlphaGenome 2 takes a different approach: it answers one question at a time — "what does this DNA sequence do in this exact cell type, in this exact state?" — and DeepMind says that focus is what let it roughly double its hit rate for finding disease-causing variants in real patient data.

What is AlphaGenome 2?

AlphaGenome 2 is a DNA sequence model that takes up to one million letters of genetic code as input and predicts molecular readouts tied to a specific tissue, cell type, or cell state. Those readouts include gene expression, histone modifications, chromatin accessibility, and — new in this version — DNA methylation, a chemical marker on DNA that DeepMind describes as a widely requested addition from the research community.

The model can also fill in missing experimental data through what DeepMind calls cross-modality imputation: predicting one type of measurement, such as chromatin accessibility, from a different one, such as single-cell gene activity, for cells where that data was never collected. The paper behind the release, authored by Adam Gayoso, Natasha Latysheva, Pushmeet Kohli, and Žiga Avsec, is published as a technical report alongside the blog post.

What changed from the original AlphaGenome

The core difference is architectural. The first AlphaGenome predicted outcomes across many tissues in parallel from a single pass of the model, which DeepMind now says "spread its capacity thin" across cell types it was never optimized for. AlphaGenome 2 is a conditional model: it generates a prediction for one named condition — a tissue, a cell type, or a cell state — using that condition as an input alongside the DNA sequence itself.

That conditioning also works in reverse. A built-in cell encoder lets researchers feed in their own single-cell sequencing data to define a custom condition, rather than being limited to the cell types DeepMind trained on. DeepMind says this is what allows the model to generalize to brand-new datasets without retraining — a claim it backs with a case study on a COVID-19 vaccine trial, described further below.

AlphaGenome 2 at a glance
  • Announced: November 11, 2026, by Google DeepMind
  • Input: DNA sequences up to 1 million base pairs long
  • New readout: DNA methylation, added to gene expression, histone marks, and chromatin accessibility
  • Training data: More than 107 million individual cell profiles across 68 human and mouse tissues
  • Access: Free API for non-commercial research now; commercial access via a Google Cloud private preview
  • Use: Research only — not validated for clinical or diagnostic use

How the model reads a genome

AlphaGenome 2 scores the effect of a genetic variant by running the model twice on the same stretch of DNA — once with the reference sequence and once with the mutation inserted — and comparing the two predicted outputs for the chosen cell type or state. The bigger the gap between the two predictions, the more likely DeepMind's system is to flag that variant as functionally important for that specific cell.

Because the model is conditional rather than multi-tissue, each prediction is tailored to the biological context a researcher specifies, rather than being an average or a best guess across unrelated cell types. DeepMind frames this as closer to how a cell biologist actually works: asking what a mutation does in, say, a resting T cell versus the same mutation in an activated T cell, rather than asking what it does to "the genome" in general.

The training data behind the jump

DeepMind says AlphaGenome 2 was trained on bulk and single-cell sequencing data from human and mouse samples, totaling more than 107 million individual cell profiles spanning 68 tissues. That is a far larger and more granular dataset than what powered the original model, and DeepMind credits it directly for the system's ability to distinguish between closely related cell states that look nearly identical in bulk sequencing data.

DeepMind is continuing to expand that pool through an ongoing collaboration with the Human Cell Atlas consortium, the international project mapping every cell type in the human body, to profile additional cell types, developmental lineages, and experimentally perturbed cell states.

The performance numbers DeepMind is citing

DeepMind's headline comparison uses the TenK10K dataset, which covers 28 primary human immune cell types, to test how well each model finds variants that actually change gene expression in real single-cell data. The lab says AlphaGenome 2 roughly doubles the hit rate of the original AlphaGenome and other specialized single-cell tools on that benchmark.

MetricAlphaGenome (original)AlphaGenome 2
Public releaseJune 2025 (peer-reviewed in Nature, January 2026)November 11, 2026
Prediction styleMultiple tissues predicted in parallelConditional — one tissue/cell type/state at a time
Core readoutsGene expression, histone marks, chromatin accessibility, splicingSame, plus DNA methylation
Training scaleSmaller multi-tissue reference set107M+ single-cell profiles across 68 human and mouse tissues
Causal variants found (TenK10K, 28 immune cell types)16%30%
Correct direction of gene effect23%43%
Custom cell types from new dataRequires retrainingCell encoder adapts without retraining
Commercial accessNon-commercial API onlyNon-commercial API now; commercial private preview via Google Cloud

DeepMind also says the new model explains more of the variation in gene expression between cell types in genomic regions it was never trained on, and that its predictions line up more closely with observed expression levels during hematopoiesis, the process by which blood cells form and specialize.

Two case studies DeepMind points to

To demonstrate the model outside of benchmark scores, DeepMind highlights two examples. The first reproduces a previously published finding in inflammatory bowel disease: a specific genetic variant that increases activity of the ETS2 gene in gut macrophages, immune cells implicated in IBD. DeepMind says AlphaGenome 2 recovers that same effect directly from sequence, without being trained on IBD data specifically.

The second example tests the model's claim that it generalizes to new datasets without retraining. DeepMind applied AlphaGenome 2, unmodified, to data from a COVID-19 vaccine response study and found it could distinguish activated T cells from resting T cells — a distinction the model was never explicitly trained to make — matching or beating models that were built and trained specifically for that dataset.

How to access AlphaGenome 2

Non-commercial researchers can use AlphaGenome 2 today through the AlphaGenome API documentation, which includes interactive Colab notebooks for setting up expression-conditioning queries. DeepMind says precomputed predictions are also available for trait-associated variants across hundreds of cell and tissue conditions, so researchers don't need to run the model themselves for common use cases.

Commercial users can request access through a private preview on Google Cloud; DeepMind has not published pricing for that tier. The lab also says AlphaGenome 2 "Skills" are coming to Google Antigravity, its agentic development environment, though it hasn't given a date. Separately, DeepMind is running an open call for research use cases through November 30, 2026, with selected projects receiving additional API quota and compute credits.

Early reaction from genomics researchers

DeepMind's announcement includes comments from two outside researchers already working with the model. Professor Sarah Teichmann, a co-founder of the Human Cell Atlas consortium at the University of Cambridge, said the system "adapts directly to new single-cell data without retraining" — pointing to the same generalization behavior demonstrated in the COVID-19 vaccine case study.

Peter Donnelly, chief executive of Genomics Limited, said his company is looking at incorporating AlphaGenome 2 into Mystra, its genomic analysis platform. Neither researcher's comments were independently verified beyond DeepMind's own post, and both are described by DeepMind as partners already engaged with the model ahead of its public release.

Why a cell-specific genome model matters

Most of the roughly 90% of the human genome that doesn't code for proteins still plays a role in switching genes on and off, but figuring out exactly what a given stretch of that non-coding DNA does — and in which cell type — has been one of the hardest problems in genomics. A genetic variant might be harmless in most cells but disease-causing in one specific cell type, which is exactly the kind of signal that averaged, multi-tissue models tend to wash out.

AlphaGenome 2's conditional design is DeepMind's attempt to close that gap, and it fits a broader pattern at the lab of releasing narrower, more specialized models rather than one model that tries to do everything. DeepMind has taken a similar approach elsewhere this year with Gemini Robotics 2.5, a model built specifically for physical robot control rather than general reasoning, while competitors such as Nvidia have pursued their own narrow, task-specific systems like GR00T for humanoid robots. AlphaGenome 2 sits alongside DeepMind's more general-purpose work, including the Gemini 4 Carbon flagship model, as evidence that the lab is pairing its broad frontier models with deep, domain-specific ones for science.

What the model can't do yet

DeepMind is explicit that AlphaGenome 2's predictions are for research use only and have not been validated or approved for clinical or diagnostic purposes — the same caveat it attached to the original AlphaGenome. The lab's own genomics lead has also said the new model won't immediately replace the first version: AlphaGenome 2's detailed, cell-type-specific predictions are more computationally expensive to run, which makes it impractical to use for the kind of genome-wide screening DeepMind already built with the original model.

That earlier work is the AlphaGenome Atlas, a roughly 1-petabyte database DeepMind and Google published on September 8, 2026, that pre-calculates the predicted regulatory effect of all 9 billion possible single-letter changes across the human genome, searchable through a no-code web portal. Running AlphaGenome 2's conditional, cell-type-aware predictions at that same genome-wide scale isn't yet practical, so for now the two systems are complementary: the Atlas for broad, every-variant screening, and AlphaGenome 2 for the smaller number of variants where knowing the exact cell type matters most.

What to watch next

The near-term dates to track are DeepMind's November 30, 2026 deadline for researchers to submit use cases for extra API quota and compute credits, and the still-unscheduled rollout of AlphaGenome 2 "Skills" inside Google Antigravity. Longer term, the real test will be whether outside labs — including Genomics Limited and the Human Cell Atlas researchers already cited by DeepMind — can independently reproduce the model's TenK10K results and its COVID-19 T-cell case study, since all of the performance figures published so far come from DeepMind itself rather than third-party replication. A commercial pricing announcement for the Google Cloud preview is also still pending.

Frequently asked questions

What is AlphaGenome 2?

AlphaGenome 2 is an AI model from Google DeepMind that predicts how a DNA sequence affects gene expression, histone modifications, chromatin accessibility, and DNA methylation in one specific cell type or state at a time. It was announced on November 11, 2026.

How is AlphaGenome 2 different from the original AlphaGenome?

The original AlphaGenome predicted outcomes across many tissues in parallel. AlphaGenome 2 is a conditional model that generates a prediction for one named cell type or state at a time, and DeepMind says this roughly doubles its hit rate for finding disease-causing genetic variants in benchmark tests.

How much better is AlphaGenome 2 at finding disease variants?

On the TenK10K dataset covering 28 primary human immune cell types, DeepMind says AlphaGenome 2 correctly identified causal variants 30% of the time, versus 16% for the original AlphaGenome and other specialized tools. Accuracy on the direction of a gene's effect rose from 23% to 43%.

Can researchers use AlphaGenome 2 right now?

Yes. Non-commercial researchers can access AlphaGenome 2 now through the AlphaGenome API, with interactive Colab tutorials. Commercial users can request access through a private preview on Google Cloud; DeepMind has not published commercial pricing.

Is AlphaGenome 2 approved for clinical or diagnostic use?

No. DeepMind says AlphaGenome 2's predictions are intended for research use only and have not been validated or approved for clinical or diagnostic purposes.

Does AlphaGenome 2 replace the original AlphaGenome or the AlphaGenome Atlas?

Not immediately. DeepMind says AlphaGenome 2's cell-type-specific predictions are more computationally expensive, making it impractical to use at the genome-wide scale of the AlphaGenome Atlas, a roughly 1-petabyte database covering all 9 billion possible single-letter DNA changes. The two tools are meant to complement each other.

Sources

More on AlphaGenome 2 →AlphaGenome 2Google DeepMindAI in genomicsDNA variant predictionGoogle AI
Theo Park
Written byTheo Park

Theo Park runs the AI desk at Pandromeda. He follows model launches from the frontier labs and the open-weight community, tracks the assistants and developer tools built on them, and explains what each release changes on pricing, capability and safety. His reporting leans on primary sources: model cards, technical reports, API documentation and the companies' own announcements.

More from AI

See all