What is AlphaGenome 2 — DeepMind's DNA model now reads variants cell by cell, 30% recall versus 16%

Google DeepMind released AlphaGenome 2 on November 11, 2026. Where the first AlphaGenome, from June 2025, could say that a DNA variant alters gene activity in liver tissue, the second tries to say in which liver cells. DeepMind says the model identifies nearly twice as many variants that change gene expression as the previous best model. Nature, reporting the same day, quoted a researcher who said the model reached in 30 minutes part of an answer that had taken his lab about a year and a half of experiments.

It is a research tool. It does not diagnose disease from a personal genome, DeepMind states that it has not been validated or approved for clinical or diagnostic purposes, and every performance figure so far is the company's own.

1. What it does — a million DNA letters, plus which cell you mean

A few terms first.

Most DNA variants linked to disease lie outside protein-coding genes. They do not break a protein; they nudge a switch. The difficulty is that every cell carries the same DNA but uses different switches. A switch that matters only in gut immune cells is lost in a tissue-wide average.

Nature reports that the training data span 637 human and 311 mouse cell types.

DeepMind calls the design a conditional model. Version 1 took DNA and returned a fixed set of predictions for many tissues and cell lines in parallel. DeepMind says that spread the model's capacity thin and made closely related cell types hard to tell apart. Version 2 takes one condition at a time and feeds it into every stage of the network, so the same stretch of DNA is read differently depending on the cell.

Two features follow from that. Expression conditioning lets a researcher upload single-cell RNA data from an experiment — a few dozen cells are said to be enough — which the model compresses into a cell-state representation and uses as the condition, with no retraining. That extends predictions to cell states the model never saw, such as cells from a disease sample. Cross-modality imputation means that given one measurement, such as single-cell gene activity, the model predicts properties that were not measured, such as chromatin accessibility, which helps decide which experiment to run next. Sarah Teichmann, the University of Cambridge professor who co-founded the Human Cell Atlas consortium, said in a comment published by DeepMind that this adaptation to new single-cell data without retraining is what makes the model exciting.

2. How it differs from AlphaGenome 1

The first AlphaGenome launched on June 25, 2025, and its paper appeared in Nature on January 28, 2026. In September 2026 DeepMind added the AlphaGenome Atlas, precomputed predictions for all roughly 9 billion possible single-letter changes in the human genome. Version 2 is a separate model, not an extension of the Atlas.

The headline comparison comes from TenK10K, a single-cell dataset covering 28 primary human immune cell types. The baseline is not the stock first model but AlphaGenome 1 fine-tuned on single-cell data, along with other specialised tools.

The first two rows appear in DeepMind's post and across coverage; the third row and the 90% threshold come from reports citing the paper. "Nearly twice" is 30 against 16.

Three cautions apply. Thirty percent is not an accuracy score: it is the share of causal variants recovered under a strict bar against false positives, which also means roughly 70% were not recovered at that bar. The figures come from one immune dataset, with no matching numbers published for other tissues. And they are self-reported.

The showcase: inflammatory bowel disease and ETS2. A 2024 Nature study led by James Lee at the Francis Crick Institute in London found that a variant in a "gene desert," far from any protein-coding sequence, over-activates the gene ETS2 in immune cells and raises the risk of inflammatory bowel disease. DeepMind says AlphaGenome 2 picked out the causal variant among the candidates and predicted that it boosts ETS2 selectively in gut macrophages, leaving other cell types and genes largely untouched. Reports citing the paper put the candidate set at seven variants and the distance to the gene at about 290,000 letters.

According to Nature, Lee said the model reached its answer in 30 minutes where his team needed about a year and a half of lab work, and called that "reassuring from our perspective." He also said a model unable to predict cell-type specificity had not been much use to his group, calling that gap in the first version a "bit of a problem." This is a replay of a known result, not a new mechanism discovered by the model.

A second example used single-cell data from a COVID-19 mRNA vaccine study that the model had not seen. DeepMind says the model separated vaccine-activated T cells from resting ones and predicted that a variant linked to severe COVID-19 disrupts gene activity more strongly in the activated state.

3. Who can use it, and how

For a non-commercial researcher, the steps in the API client repository are short: get a key at alphagenome.google/api; run pip install alphagenome; connect through the version 2 client module (alphagenome.models.v2); then request predictions for a one-million-letter interval under a chosen condition, or score a variant. To use your own single-cell RNA data as the condition, DeepMind has published Colab tutorials on expression conditioning.

The same repository sets limits. Live predictions suit small-scale analyses, not genome-wide sweeps; lookups in the precomputed Atlas get higher query rates. Model outputs and Atlas information are for non-commercial use and must not be used to train other machine learning models. Predictions are for theoretical modelling and research, and must not be used for clinical decision-making or relied on as medical advice.

On weights, version 1 sets a precedent without a promise. It began as an API preview; with the Nature paper in January 2026 the weights went up on Kaggle and Hugging Face under non-commercial model terms, and the research repository recommends at least an Nvidia H100 GPU to run them. For version 2 the post says nothing about weights, and the version 1 repository has no entry for it.

Commercial terms are described differently in two places. The DeepMind post calls commercial research access to version 2 a private preview; the API client README says commercial use is available through Google Cloud, without saying which model that covers. Peter Donnelly, chief executive of Genomics Limited, said in the post that the company looks forward to incorporating AlphaGenome 2 into its Mystra platform. DeepMind is also taking submissions for a use-case spotlight until November 30, with extra quota and credits possible for selected projects.

4. AlphaFold, AlphaMissense, AlphaGenome — which does what

The names blur together, but the models sit at different stages. DNA is read into RNA, RNA is translated into protein, and the protein folds and does its job. AlphaGenome addresses the first stage; AlphaFold addresses the last.

AlphaMissense and AlphaGenome both score variants, but on different ground. AlphaMissense, adapted from AlphaFold, scored all 71 million possible missense variants — single-letter changes that alter one amino acid — and classified about 89% as likely pathogenic or likely benign. Protein-coding DNA is about 2% of the genome. AlphaGenome takes the other 98%, where variants change how genes are switched rather than what proteins are made. DeepMind described the two as complementary when it launched version 1.

AlphaFold led to a share of the 2024 Nobel Prize in Chemistry for Demis Hassabis and John Jumper and a public database of roughly 200 million structures. The AlphaGenome line is not at that stage. Protein structures have experimentally determined answers to grade against; gene regulation changes with cell, time and environment, and ground-truth data are scarce.

5. Limits and what is not confirmed

It is not a diagnostic. DeepMind states that both AlphaGenome versions are for research use only. With version 1 it also said the model was not designed or validated for personal genome prediction, and that it predicts molecular outcomes without explaining complex traits or diseases shaped by development and environment.

Limitations the paper itself lists, as relayed in coverage:

Cost is why version 1 stays. Žiga Avsec, who leads the AlphaGenome team, is quoted as telling The Verge that the new predictions are "more expensive," which made a genome-wide atlas built on version 2 impractical. AlphaGenome 1 and the Atlas remain in service.

Published reaction so far is favourable, with caveats about who is speaking. The outside scientist quoted by Nature is the author of the study that was reproduced. The commentators in DeepMind's post were chosen by DeepMind. The paper is a preprint. The showcase is a result already known. What does not yet exist is a benchmark run by a group with no stake, a critical outside review of version 2, a case in which the model predicted first and an experiment confirmed afterwards, a commercial price, or a decision on weights. Some coverage also notes the dependence that API-only access creates: researchers cannot inspect or self-host the model, and the ban on training other models with its outputs constrains follow-on work.

DeepMind says it will add more human cell types, developmental lineages and perturbed states, including through work with the Wellcome Sanger Institute, and that agent tooling is next; Google's agent models are covered in our "Gemini 4 Carbon launch". For another recent case of DeepMind pointing AI at a research problem, see "DeepMind AI agents find two RSA-1024 weaknesses".