DEV Community

Cover image for AlphaGenome 2 vs AlphaGenome 1: Cell-Type DNA Effects
Shamuddin
Shamuddin

Posted on Originally published at aitecharchive.com

AlphaGenome 2 vs AlphaGenome 1: Cell-Type DNA Effects

AlphaGenome 2 is the better tool for any question about which cell type a DNA variant acts in, and AlphaGenome 1 is now mainly a reference point. On 11 November 2026, Google DeepMind announced the model, which takes a DNA sequence plus a chosen cell type, tissue or cell state and predicts what a variant does there. On TenK10K, a single-cell map of 28 primary immune cell types, DeepMind reports that it identified 30% of causal variants against 16% for the original model fine-tuned on single-cell data, and 43% of directional gene effects against 23%.

Last verified: 13 November 2026. The model is for research use only and is not validated or approved for clinical or diagnostic purposes.

TL;DR

  • AlphaGenome 2 is a conditional model: one 1-million-letter DNA sequence plus one condition (tissue, cell type or cell state) per prediction.
  • On TenK10K, DeepMind reports 30% vs 16% causal variants (DeepMind announcement) and 43% vs 23% directional gene effects (preprint), against the original model fine-tuned on single-cell data and other specialised tools.
  • It adds DNA methylation as an output and can impute unmeasured properties, such as chromatin accessibility, from one measurement.
  • Access is through the AlphaGenome API for non-commercial research, with Colab tutorials; commercial research is in private preview.
  • The technical claims come from a preprint that has not been peer reviewed, and the benchmark was run by the model's developers.

What is AlphaGenome 2 and what changed from the first model?

AlphaGenome 2 is a DNA-sequence model built to predict how sequences and their variations affect specific human cell types, with the stated aim of finding the cellular drivers of disease. The first AlphaGenome, launched in 2025, read a 1-million-letter sequence and predicted properties across thousands of experiments from many tissues in parallel. That gave a broad picture but blurred the cell type in which a variant matters.

The new version changes the architecture rather than just the scale. It is conditional: you supply the sequence and one condition at a time, and that condition modulates the sequence representations at every stage of the network. The question moves from "what does this variant do somewhere in the body" to "what does it do in this cell type, in this state".

This also addresses a gap in the AlphaGenome Atlas released about two months earlier, in September 2026. The Atlas carried predictions for all 9 billion possible single-letter DNA changes, but, as Nature noted, those predictions did not specify the cell type in which a variant acts.

How does AlphaGenome 2 compare with AlphaGenome 1?

The table summarises what DeepMind and Nature report. Where the announcement does not state a figure for the first model, the cell says so rather than guessing.

Feature AlphaGenome 1 AlphaGenome 2
Conditioning None; predicts many tissues in parallel One condition per run: tissue, cell type or cell state
Training data Not restated in the announcement Bulk and single-cell data, human and mouse; over 107 million cell profiles across 68 tissues
Outputs Gene expression, histone modifications, chromatin accessibility Same, plus DNA methylation
TenK10K causal variants 16% (fine-tuned on single-cell data) 30%
TenK10K directional gene effects 23% 43%
Access See the original release API for non-commercial research; commercial in private preview

Nature reports the model was trained on 637 human and 311 mouse cell types, though users can supply new single-cell gene-activity data to get predictions for states outside that catalogue. The preprint abstract adds that the model doubles the recall of causal single-cell eQTLs at 90% precision (preprint) and generated an atlas of enhancer-gene links and scores for 7 million common variants (preprint) across hundreds of human cell types. Treat those as preprint claims until peer review.

Does AlphaGenome 2 reproduce known disease biology?

Yes, in at least two examples DeepMind describes. The first is a 2024 Nature study showing that a variant in a distant regulatory switch drives inflammatory bowel disease by overactivating the ETS2 gene in gut macrophages. AlphaGenome 2 singled out the causal variant and showed it selectively boosted ETS2 in gut macrophages, largely leaving other cell types and genes unaffected.

The second uses unseen single-cell data from a COVID-19 mRNA vaccine study. The model separated vaccine-activated T cells from resting ones and showed that a variant linked to severe COVID-19 disrupts gene activity more strongly in the activated state, an effect missed in unactivated cells. Both are retrospective checks against known or newly supplied data, not prospective discoveries, and the first is a reproduction of a published result.

How do you use AlphaGenome 2 on your own cell data?

DeepMind describes a workflow that needs no retraining. A researcher uploads a single-cell RNA-seq snapshot, which can be as small as a few dozen cells. The data is compressed into a cell-state representation that conditions the model on a state it never saw in training. DeepMind reports this can match or outperform models trained or fine-tuned directly on those datasets, which is a vendor claim worth testing on your own data.

The same mechanism supports cross-modality imputation. Given one measurement, such as single-cell gene activity, the model can predict properties you did not measure, such as chromatin accessibility. In practice that helps decide which experiment to run next.

How to get started

  1. Request access to the AlphaGenome API through DeepMind's announcement page; non-commercial research use is available now.
  2. Work through the interactive Colab tutorials on expression conditioning to see the input format.
  3. Pick a condition: a catalogued cell type or tissue, or a cell-state representation built from your own single-cell RNA-seq data.
  4. Score variants of interest and check the predicted direction (up or down) where the model is confident.
  5. Use precomputed scores for trait-associated variants, covering traits from height to type 2 diabetes, as a sanity check before running your own.

DeepMind also says AlphaGenome 2 Skills are coming to Google Antigravity, and it asked researchers to submit use cases by 30 November 2026 for a chance at extra quota and credits.

What are the limits of AlphaGenome 2?

There are four to keep in mind. First, the headline numbers are from DeepMind's own preprint, which has not been peer reviewed, so independent replication matters. Second, the 30% and 16% figures describe one benchmark on 28 immune cell types, and performance on other tissues may differ. Third, the model is research only and carries no clinical or diagnostic validation, so it must not inform patient decisions. Fourth, a computational prediction of a causal variant still needs experimental confirmation.

Independent reaction has been positive. In Nature's coverage, Kyle Lee of the Francis Crick Institute said cell-type specificity was the missing piece that had kept earlier versions from being useful for his work.

What is the verdict for genomics and AI-for-science watchers?

For variant interpretation, AlphaGenome 2 wins. Anyone working on immune, gut or other cell-type-specific disease questions gets a model that answers the question they actually ask, and nearly doubling causal-variant recovery on TenK10K is a meaningful step if it replicates. AlphaGenome 1 remains reasonable for broad tissue-level scans, but there is little reason to start new cell-type work on it.

For wider AI-for-science watchers, the pattern is conditioning a large model on context rather than training one model per dataset. It sits alongside other recent work we have covered, including how GPT-6 Astra and Hydrogen 1.0 compare on science benchmarks, BioBasicsBench for wet-lab reasoning, an offline pathogen-screening model that runs on a laptop, and a comparison of bacterial and human vault particle structures. DeepMind's wider research agenda also appears in its work on AI agents and 1024-bit RSA.

FAQ

Q: What is AlphaGenome 2?
A: It is a DeepMind DNA-sequence model that predicts how DNA sequences and variants affect specific human cell types, using a chosen cell type, tissue or cell state as an input condition.

Q: How much better is AlphaGenome 2 than the original model?
A: On the TenK10K single-cell benchmark, DeepMind reports 30% of causal variants identified against 16%, and 43% of directional gene effects against 23%, compared with the original model fine-tuned on single-cell data. These figures come from a preprint.

Q: Can I use AlphaGenome 2 for clinical or diagnostic decisions?
A: No. DeepMind states it is for research use only and is not validated or approved for clinical or diagnostic purposes.

Q: How do I access AlphaGenome 2?
A: Non-commercial research use is available through the AlphaGenome API, with Colab tutorials on expression conditioning. Commercial research is in private preview.

Q: Does AlphaGenome 2 need retraining for new cell types?
A: No. You can upload a single-cell RNA-seq snapshot, even a few dozen cells, and the model converts it into a cell-state representation used to condition predictions.

Q: What new outputs does AlphaGenome 2 add?
A: It adds DNA methylation to the existing outputs of gene expression, histone modifications and chromatin accessibility, and can impute unmeasured properties from a single measurement.

Sources

Corrections log: none at publication. This article was produced with AI assistance and checked against the primary sources above.

Top comments (0)