AlphaGenome 2: DeepMind teaches its DNA model to ask "in which cell?"

While most of the AI industry spent this week watching Nvidia’s boardroom, Google DeepMind shipped something with a longer half-life. On Wednesday, 11 November 2026, the lab announced AlphaGenome 2, the successor to the DNA model it launched in 2025, together with a 78-page technical report. It is available from today through the AlphaGenome API for non-commercial research, with a private preview for commercial users.

The one-line pitch: the first AlphaGenome told you what a DNA variant does in a tissue, on average. AlphaGenome 2 tries to tell you what it does in one particular kind of cell, in one particular state, including cells the model was never trained on.

The problem it attacks

Almost every cell in your body carries the same genome, yet a heart muscle cell, a skin cell and an activated T cell switch on very different genes. Most variants linked to disease by genome-wide association studies (GWAS) sit in non-coding DNA: not in genes themselves but in the regulatory “switches” (enhancers, promoters) that decide when genes turn on. Whether a variant matters often depends on which cell is reading that switch.

Single-cell atlases such as the Human Cell Atlas and CELLxGENE now catalogue tens of millions of individual cells, but, as the report notes, they rarely profile enough people to link a variant to a change in expression statistically. A model that predicts variant effects straight from DNA sequence could fill that gap, if it can tell closely related cell types apart. Previous models, including AlphaGenome 1 and the academic Borzoi, were weak at exactly that.

What is actually new: conditioning, not more output heads

The architecture change is the part a machine-learning engineer will recognise. AlphaGenome 1 took a 1-megabase DNA window and predicted thousands of experimental “tracks” in parallel, one fixed output head per tissue or cell line. DeepMind’s blog concedes this “spread the model’s capacity thin”.

AlphaGenome 2 keeps the 1-Mb input and the U-Net-style backbone (convolutional encoder, transformer tower, decoder) but becomes a conditional model. It takes three inputs: the DNA sequence, a condition (a tissue, cell type or cell state) and a species flag (human or mouse). The condition is an embedding that modulates the network through adaptive normalisation layers from the earliest layers onward, so the model computes “how does this cell read this DNA” instead of emitting every answer at once.

The ablations in the report are instructive: injecting the condition only at the decoder or output barely helped; conditioning the encoder and transformer is where the gains came from. That is a familiar lesson from conditional image and audio models, now showing up in genomics.

Other details from the report:

  • Data: 3,977 single-cell RNA-seq and ATAC-seq pseudobulk tracks from 36 human and mouse studies, cell-type-resolved DNA methylation for 185 human conditions (methylation is a new output), plus the bulk data AlphaGenome 1 used.
  • Training in four stages: pre-training, condition refinement, then distillation. Eight independently trained “teacher” models are averaged and a single “student” learns to reproduce the ensemble on randomly mutated sequences, which is how DeepMind packs ensemble-level robustness into one model. Finally comes a post-training stage.
  • A cell encoder for unseen cells: in post-training, an autoencoder trained on about 107 million human single-cell profiles is mapped into the frozen model’s condition space through a lightweight adapter. A scientist can upload single-cell RNA data from “just a few dozen cells” and get a condition embedding for a cell state the model has never seen, with no fine-tuning.
  • Imputation: because conditions are decoupled from output heads, measuring one modality (say, gene expression) lets the model predict unmeasured ones such as chromatin accessibility.

The numbers

DeepMind’s headline is “nearly twice as many” variants found. In the report that means: on TenK10K, a single-cell eQTL dataset covering 28 immune cell types, recall of causal variants at 90% precision rose from 16% to 30% against AlphaGenome 1 fine-tuned on single-cell data, and correct effect direction rose from 23% to 43%. On GTEx bulk tissues the gains are smaller: recall at 90% precision went from 17% to 26%, and the overall AUPRC went from 0.72 to 0.74 (Borzoi: 0.69). For cross-tissue expression, the median Pearson correlation rose from 0.61 to 0.71.

Two worked examples carry the narrative. The model computationally recovered a published finding that a distant regulatory variant drives inflammatory bowel disease by over-activating the ETS2 gene specifically in gut macrophages. And given unseen data from a COVID-19 mRNA vaccine study, it found that a variant linked to severe COVID disrupts gene activity more strongly in activated T cells than in resting ones, an effect a tissue-average model would miss.

DeepMind is also releasing precomputed resources: enhancer-to-gene links across 640 conditions and scores for nearly 7 million trait-associated and common variants.

Why it matters beyond biology

For readers of an IT blog, three points stand out.

  1. Conditioning beats multi-head. Moving from “predict everything in parallel” to “predict for the context you specify” is the same design shift that made text-to-image and instruction-following models work. AlphaGenome 2 is evidence that it generalises to scientific foundation models.
  2. Adapters as an extension point. The frozen backbone plus a small adapter that maps a new data type into the condition space is a clean software pattern. The report already suggests adapters for spatial transcriptomics and even natural-language descriptions of cells.
  3. Models as tools for agents. DeepMind says “AlphaGenome 2 Skills” are coming soon to Google Antigravity, its agentic development platform, so that an agent like Gemini 4 can call the model during a research workflow. Specialist models becoming callable skills for general agents is a pattern worth watching well outside genomics.

The skepticism section

The report itself is unusually candid. “At high precision, AlphaGenome 2 recovers only a minority of fine-mapped eQTLs and putatively causal GWAS variants”: 30% recall means 70% of known causal variants are still missed at that precision. The model tends to underestimate effects outside accessible chromatin, at very distal enhancers, or through post-transcriptional mechanisms. It still struggles with fine-grained states such as neuronal subpopulations. And because it learns from observational, steady-state data, its condition embeddings are correlational. Whether it can predict what happens after a gene knockout or during a state transition is, in DeepMind’s words, “unclear”.

Cost is the other caveat. Each condition needs its own forward pass, so AlphaGenome 2 is more expensive than its predecessor. Genomics lead Žiga Avsec told The Verge that this is why it could not have been used to build September’s genome-wide AlphaGenome Atlas, and why it complements rather than replaces AlphaGenome 1.

Finally, every benchmark comes from DeepMind’s own paper, which is a preprint-style technical report, not a peer-reviewed publication (AlphaGenome 1 reached Nature only months after launch). The tools are labelled research-only and not validated for clinical or diagnostic use. Independent replication on the API is the real test, and DeepMind is courting exactly that, offering extra quota for community projects submitted by 30 November.

The direction is clear even if the recall is not yet impressive. Long-term, Avsec frames models like this as building blocks for “virtual cells”. For now, AlphaGenome 2 is a sharper filter that tells a biologist which variant, in which cell, to put on the bench first.

Sources