Research review / research

AlphaGenome 2 Predicts DNA Effects One Cell Type at a Time, and Nearly Doubles Variant Recall on Immune Cells

Google DeepMind released AlphaGenome 2 on November 11. Instead of predicting every tissue in parallel, it takes one cell type or state as input alongside a million-letter DNA sequence. Trained on more than 107 million cell profiles, it found 30% of causal variants on a 28-cell-type immune benchmark against 16% for the earlier model. Access is through an API, and it is for research only.

By Sandeep Belgavi · Published 2026-11-14

On November 11, 2026, Google DeepMind announced AlphaGenome 2, "our next-generation model that predicts how DNA sequences and their variations impact specific human cell types." It is "available today for non-commercial research via the AlphaGenome API, and in private preview for commercial research." A paper accompanies the post.

What changed in the design

The first AlphaGenome "took a million-letter DNA sequence and predicted properties across thousands of experiments from many tissues all at once." DeepMind says that approach "spread the model's capacity thin and made it hard to capture subtle differences between closely related cell types." AlphaGenome 2 "is a conditional model that predicts DNA effects for one condition - defined as a tissue, cell type or state - at a time," and uses that condition "as input at every stage of its architecture."

Two consequences follow. The model can fill in missing measurements: "if a scientist inputs just one measurement, like single-cell gene activity, AlphaGenome 2 can then predict unmeasured properties, such as chromatin accessibility." And it can work on new data without retraining. Scientists "can upload a single-cell RNA-seq snapshot of gene activity from just a few dozen cells", which a cell encoder compresses into a state the model conditions on.

It was trained on "bulk and single-cell data from human and mouse samples, including over 107 million individual cell profiles", and adds DNA methylation to the readouts the first model predicted.

The reported results

  • On TenK10K, "a large single-cell map of 28 primary immune cell types", it "identified nearly twice as many causal variants (30% vs. 16%) and directional gene effects (43% vs. 23%)" compared with the original model fine-tuned on single-cell data and other tools.
  • It reproduced, computationally, a published finding that a distant regulatory variant drives inflammatory bowel disease by over-activating the ETS2 gene in gut macrophages.
  • On unseen data from a COVID-19 mRNA vaccine study, it separated activated T cells from resting ones and found a variant whose effect is stronger in the activated state.

Why it matters beyond biology

The architectural change is a familiar trade-off. A model that predicts everything at once is cheap to run but blurs fine distinctions. A model that takes the context as an input can specialise per query and handle contexts it never saw, at the cost of one pass per condition. DeepMind also precomputed predictions for trait-associated variants across hundreds of cell and tissue conditions, so most users can look results up without running the model.

What remains uncertain. All figures are DeepMind's. The benchmark comparison is against its own earlier model and selected tools. The post does not mention weights or license terms beyond API access. DeepMind says the models "are intended for research use only and have not been validated or approved for clinical or diagnostic purposes."

Analysis: The conditioning idea transfers to ordinary products. At Belgavi's Book Store, Tanishka's recommendation model currently predicts interest across all reader segments in one output. A version that takes the segment, such as a school reading list or a returning collector, as an input could be tested on a segment it was not trained on, using a few dozen recent orders as the condition, which is the same trick at much smaller scale.

Sources and evidence

AlphaGenome 2, released November 11, conditions on one tissue, cell type or state per prediction instead of predicting all tissues in parallel. Trained on over 107 million cell profiles, it reports 30% causal-variant recall against 16% for its predecessor on 28 immune cell types, imputes missing measurements and adapts to new single-cell data without retraining. API access, research use only.

Views are my own and do not represent any employer, past or present.