Google DeepMind has just reset the field of computational biology, and AlphaGenome 2 is the reason why. The model reads DNA the way a living cell does, one cell type at a time, and it is already rediscovering the genetic roots of disease.
The genome was never meant to be read in bulk
The human genome is often described as a book, but that metaphor hides the hard part. Every cell in the body carries the same three billion letters, yet a neuron, a gut macrophage and a red blood cell each read that text completely differently. The same mutation can be silent in one cell type and catastrophic in another. For decades, that context sensitivity has been the wall between reading DNA and understanding disease, and it is the wall that AlphaGenome 2, Google DeepMind's new genomics model, was built to break.
Google DeepMind's original AlphaGenome, released last year, took a swing at the wall. It ingested roughly a million letters of DNA at a time and predicted thousands of experimental readouts across many tissues at once. The approach was powerful, but it had a structural flaw the company now admits openly. By producing fixed, parallel outputs for bulk tissues and cell lines, the model spread its capacity thin. It could describe the genome in aggregate, but it struggled to speak the dialect of any single cell type with precision.
AlphaGenome 2, unveiled on November 11, rebuilds the system around that weakness. Instead of predicting everything everywhere all at once, it is a conditional model that focuses on one biological condition, a specific tissue, cell type or cellular state, at a time. That shift sounds modest. In practice it changes what the technology can be used for, and who gets to use it.
How AlphaGenome 2 actually works
The core idea behind AlphaGenome 2 is that the cellular context is not an afterthought bolted onto the prediction. It is baked into the model at every stage of the architecture. Alongside the roughly one million letters of sequence, the system takes a condition as a direct input, then predicts how DNA and its variations will behave inside that exact cellular setting.
Those conditions can come from a reference catalog of human cell types, or they can be generated on the fly from a researcher's own single-cell experiment. DeepMind built a dedicated cell encoder that can translate a single-cell RNA snapshot, drawn from as few as a few dozen cells, into a condition the model understands. That means a lab studying a rare tumor subtype or an obscure immune population does not have to wait for anyone to curate a dataset. It can point AlphaGenome 2 at its own cells and get predictions tuned to them, without retraining the underlying model.
The model was trained on more than 107 million individual cell profiles spanning 68 tissues, combining bulk and single-cell data from human and mouse samples, with data drawn from efforts including the Human Cell Atlas and a collaboration with the Wellcome Sanger Institute. It expands the menu of predictions too. AlphaGenome 2 keeps the earlier readouts of gene expression, histone modifications and chromatin accessibility, and adds DNA methylation, a widely requested measurement that governs how genes are switched on and off. It can also perform cross-modality imputation, taking one type of measurement as input and predicting an unmeasured one, filling gaps that would otherwise require expensive new experiments.
The numbers that matter
Ambition is cheap in AI. The question is whether AlphaGenome 2 is measurably better, and on DeepMind's own benchmarks the answer is yes by a wide margin. Across the board, the company reports the model identifies nearly twice as many variants that change gene expression in cells compared with its previous state-of-the-art system.
The headline test came on TenK10K, a large single-cell map of 28 primary immune cell types. There, AlphaGenome 2 reliably identified nearly twice as many causal variants, 30 percent versus 16 percent, and nearly twice as many directional gene effects, 43 percent versus 23 percent. The comparison was not against a straw man. The baseline bundled the original AlphaGenome fine-tuned on single-cell data together with other specialized tools, which makes the jump more striking.
Just as important is where those gains show up. DeepMind reports that on genomic regions held out of training, the model explains significantly more of the expression differences between cell types inside a single tissue, including the fine gradations as blood stem cells mature into red blood cells. That is precisely the cell-by-cell resolution that bulk models blur away, and it is where the biology of disease tends to hide.
From code to clinic
The most persuasive evidence is not a benchmark but a rediscovery. A landmark study published in Nature traced inflammatory bowel disease to a variant in a distant regulatory switch that overactivates a gene called ETS2 specifically in gut macrophages. Handed the raw problem, AlphaGenome 2 reproduced that finding computationally, singling out the causal variant and showing that it boosts ETS2 selectively in the right cell type. Recreating a hard-won experimental result in silico is the kind of validation that makes biologists pay attention.
The COVID-19 example is arguably more forward-looking. Given single-cell data it had never seen from a vaccine study, the system distinguished vaccine-activated immune T cells from resting ones, then showed that a variant tied to severe COVID-19 disrupts gene activity far more strongly in the activated state. Disease risk, in other words, can depend not just on which cell carries a mutation but on what that cell is doing at the time. DeepMind has also released precomputed predictions for trait-associated variants across hundreds of cell and tissue conditions, including for traits as different as height and type 2 diabetes, and is extending the model to map the enhancers that control genes in primary human cell types such as brain astrocytes.
Industry is already circling. Peter Donnelly, chief executive of Genomics, called AlphaGenome 2 an impressive advance and said the company looks forward to incorporating it into its Mystra drug discovery platform. DeepMind is careful to stamp the obvious caveat on all of it. The model is intended for research use only and has not been validated or approved for clinical or diagnostic purposes. The distance between a confident prediction and a treatment decision remains long, and it is measured in clinical trials, not benchmarks.
A land grab for the genome
Access is where strategy shows. AlphaGenome 2 is available now for non-commercial research through the AlphaGenome API, with a private preview for commercial research, interactive tutorials for scientists, and an open call for use cases that researchers can submit until November 30 for extra compute quota. DeepMind has also signaled that AlphaGenome 2 skills are coming to Google Antigravity, its agentic developer environment, which hints at a future where a scientist describes a hypothesis in plain language and an agent runs the genomic analysis.
Read alongside AlphaFold, the protein-structure model that earned its creators a Nobel Prize, the new system looks less like a single product and more like another floor in a tower. DeepMind is methodically building an AI stack for biology, from protein folding to regulatory genomics, and offering it through APIs that make its tools the default substrate for a generation of researchers. Sarah Teichmann, co-founder of the Human Cell Atlas consortium, praised the model precisely because it adapts directly to new single-cell data without retraining, the feature that makes it sticky inside a working lab.
There is a sober counterweight. Impressive in-house benchmarks are not the same as independent, real-world validation, and a model that predicts regulatory effects with confidence can still be wrong in ways that only a wet lab will catch. Concentrating so much foundational biology infrastructure inside one company raises its own questions about access, reproducibility and who ultimately controls the maps that drug developers will come to depend on. For now, those debates are downstream of a simpler fact. With AlphaGenome 2, the gap between sequencing a genome and understanding what its errors will do to a living cell just narrowed, and the pace of that narrowing is what should make the rest of the field nervous. For more on the companies racing to industrialize AI for science, see Blumefield, and follow our continuing coverage of the AI research frontier.
Primary sources: Google DeepMind's AlphaGenome 2 announcement and its technical paper; the AlphaGenome API documentation; and the Nature study on ETS2 and inflammatory bowel disease.