AlphaGenome 2 Predicts How DNA Variants Act in Specific Cells — Twice the Recall of Its Predecessor

Google DeepMind announced AlphaGenome 2 today, November 11, 2026 — a deep-learning model that predicts how DNA variants affect gene regulation in specific cell types and states, rather than in a tissue average. It is trained on more than 100 million single-cell profiles, and on the benchmark DeepMind highlights, it recovered 30% of putatively causal expression-linked variants at 90% precision, against 16% for AlphaGenome 1 fine-tuned on single-cell data. Nearly double the recall, at the same precision. That's the headline number, and it's worth understanding exactly what it does — and doesn't — mean.

I haven't run AlphaGenome 2 myself. This article is built from DeepMind's announcement and the technical paper it points to, as reported on November 11, 2026, and checked the same day. Where a number is DeepMind's, I'll say so.

The blind spot it's built for

Most genomic models predict what a variant does across a whole tissue. The problem is that a variant can matter only in one specialised cell type, or only in a cell activated by disease — and in a tissue average, that signal gets diluted until it disappears. A variant linked to inflammatory bowel disease might act in gut monocytes and nowhere else. A variant tied to a vaccine response might only matter in activated T cells, not resting ones. Averaging hides exactly the cells you care about.

AlphaGenome 2 is conditioned on cell identity. You pick a tissue, a cell type or a cell state — including a condition generated from your own single-cell experiment — and the model predicts how a variant behaves in that condition specifically. The same DNA segment can be read differently in different cells, and this model is built to represent that.

That is a real change from AlphaGenome 1, which predicted molecular activity across broad tissues and cell lines at once from a one-million-letter DNA sequence. AlphaGenome 2 evaluates one condition at a time and supplies that condition throughout the network. Its readouts include gene expression, RNA splicing and chromatin accessibility, with CpG DNA methylation added to what the original model could do.

What it can do with your own data

The most practical part of the design is that researchers can supply single-cell RNA-seq data — a snapshot of gene activity — from a new experiment, and a cell encoder turns it into a condition without further fine-tuning. That means predictions for datasets and cell states the model has never seen before, without retraining.

The same design can also fill in missing experimental readouts. Given gene-expression data alone, AlphaGenome 2 can predict unmeasured properties such as chromatin accessibility — whether stretches of DNA are open to regulation — which helps researchers prioritise what to test next in the lab.

The trade-off is computational. Every condition requires a separate prediction, so the model is more intensive to run than fixed-track multitask models that produce all their outputs at once. If you want predictions across hundreds of cell types, you pay for hundreds of runs.

The benchmark numbers, in plain language

On TenK10K, a single-cell immune dataset spanning 28 cell types, DeepMind reports that AlphaGenome 2 recovered 30% of putatively causal expression-linked variants at 90% precision, compared with 16% for AlphaGenome 1 fine-tuned on single-cell data. It recovered 43% of variant effect directions at 90% accuracy, versus 23%. On genome regions excluded from training, the model explained a median 34% of within-tissue, cell-type-specific gene-expression variance, against 23% for the same fine-tuned predecessor.

Translated: at the same high precision, it finds nearly twice as many of the variants that likely cause a change in gene expression, and it gets the direction of the effect right nearly twice as often. Those are meaningful improvements — and they are DeepMind's numbers, from its own paper, on its own chosen benchmark. Independent replication is what turns a launch-day benchmark into a trusted result, and that hasn't happened yet.

DeepMind also reports two research examples. In an inflammatory bowel disease analysis, AlphaGenome 2 flagged the variant rs2836882 from a seven-variant credible set, linked it to the gene ETS2 roughly 290 kilobases away, and localised the predicted effect to gut monocytes and macrophages — reproducing a mechanism previously established experimentally. Using single-cell data from a COVID-19 mRNA vaccine study that was new to the model, it distinguished vaccine-activated T cells from resting T cells and predicted a stronger effect for a variant associated with severe COVID-19 in the activated state. Both examples show the same point: a variant's predicted effect can depend on the cell type and on the cell's current state.

What's available, and to whom

DeepMind precomputed two resources for researchers who don't want to run large-scale comparisons themselves. One maps enhancer–gene links across 640 primary single-cell reference conditions and 39,323 protein-coding and long non-coding RNA genes. Another scores more than 6.8 million unique common and trait-associated single-letter variants across 510 human conditions, to nominate possible causal variants, target genes, effect directions and relevant cell states.

At launch, AlphaGenome 2 and expression-conditioning Colab tutorials are available through the AlphaGenome API for non-commercial research, with commercial research in private preview. The official documentation says non-commercial API use is free, with query rates varying by demand. DeepMind also plans AlphaGenome 2 Skills for Google Antigravity, its agentic development platform.

Note what this is not: AlphaGenome 2 is a separate model release, not an extension of the AlphaGenome Atlas, which offered predictions for all 9 billion possible single-letter changes in the human genome. And DeepMind restricts both AlphaGenome versions to research use. Neither has been validated or approved for clinical or diagnostic purposes — this is a research tool, not a medical one.

Who this is actually for

If you work in genomics or drug discovery, the practical question is whether AlphaGenome 2 shortens the path from a disease association to a testable hypothesis. A genome-wide association study hands you a region and a list of candidate variants. What you need next is which variant, which gene, and which cell. AlphaGenome 2 is built to narrow exactly that search across hundreds of conditions — and Genomics PLC's CEO Peter Donnelly said the company looks forward to incorporating it into its Mystra platform, describing sequence-prediction models as tools for moving from association signals to the information drug-development programs need.

If you're a researcher with single-cell data of your own, the free non-commercial API and the Colab tutorials make this a cheap experiment to run this week: feed in your condition, pick a variant you already suspect, and see whether the model's prediction matches what your lab data says.

For everyone else, the honest answer is that this is a research infrastructure launch, not a product you'll use directly. Its effect reaches you later, through faster drug-target identification and better interpretation of genetic studies — if the benchmarks hold up outside DeepMind's lab.

The catch, stated plainly

• It still misses most causal variants. DeepMind's own paper cautions that even at high precision, AlphaGenome 2 recovers only a minority of the expression-linked and disease-associated variants that statistical analyses nominate as likely causal. Doubling the recall from 16% to 30% is real progress — and 70% is still missed.
• It can underestimate effects in hard places. Outside accessible chromatin, at very distant regulatory elements, or through mechanisms acting after transcription, the model's predictions get weaker. Transient, stimulated and disease states remain underrepresented in training, and fine-grained neuronal subpopulations remain difficult to distinguish.
• Cell states here are correlations, not causes. The condition representations come from observational, steady-state profiles. It remains uncertain whether the model can predict responses to gene knockouts, overexpression or dynamic cell-state transitions. The hypotheses it generates still need experimental follow-up — the model narrows the search, it doesn't close it.
• Not for clinical use. Neither AlphaGenome version is validated or approved for clinical or diagnostic purposes. Anyone presenting it as a diagnostic tool is misrepresenting it.
• Every benchmark number is DeepMind's. The 30% recall, the 43% direction accuracy, the 34% variance explained — all from DeepMind's paper. The direction of improvement is plausible and the design is sound, but independent evaluation is still to come.

The bottom line

AlphaGenome 1 told you what a variant does on average. AlphaGenome 2 tells you what it does in a specific cell, in a specific state — and on DeepMind's benchmark, it finds nearly twice as many causal variants doing it. The gap between 30% and 100% is the honest measure of where this technology stands: a sharper lens for researchers, not a finished answer. If you have single-cell data and a variant you care about, the free research API makes it worth an afternoon. If you're waiting for it to change medicine, wait for the experiments it enables, not the announcement.

Are you working with single-cell data — and would a cell-specific variant prediction change which experiment you run next?

Sources: Google DeepMind announcement via Unite.AI, "Google DeepMind's AlphaGenome 2 Predicts Cell-State Variant Effects" (https://www.unite.ai/google-deepminds-alphagenome-2-predicts-cell-state-variant-effects/), 11 November 2026; SiliconANGLE, "Google DeepMind releases new advanced robotics AI model and next-generation gene prediction model" (https://siliconangle.com/2026/11/11/google-deepmind-releases-new-advanced-robotics-ai-model-and-next-generation-gene-prediction-model/), 11 November 2026. All capability and benchmark claims are DeepMind's, stated as such. No hands-on testing was performed for this article. Nothing here is sponsored — no affiliate links on this page. This breaking article is not the daily buyer-intent post. Plain link for readers: deepmind.google.