AlphaGenome 2: Why Cell-Specific DNA Predictions Matter
- AI
- Guides
- AlphaGenome
- Genetics
- AI research
- Written by
- Bhargav Tibadiya
- Published
- Reading time
- 9 min read
A genetic variant can matter in one cell type and do very little in another. That is the central problem AlphaGenome 2 tries to address. Instead of asking only what a stretch of DNA does across broad tissue measurements, the new model also takes a biological condition, such as a cell type or state, and predicts what happens there.
Google DeepMind announced AlphaGenome 2 on November 11, 2026. It reports better cell-specific predictions and better recovery of gene-regulating variants than the previous AlphaGenome model. The practical change is finer biological context, not a claim that AI can now diagnose disease from a sequence[1].
This is a source-based explanation, not an independent model benchmark or a medical recommendation. The model is for research only. DeepMind says neither AlphaGenome nor AlphaGenome 2 has been validated or approved for clinical or diagnostic use[1].
Why the cell type belongs in the question
Most cells carry the same DNA, but they do not use every part of it in the same way. A regulatory sequence that helps control a gene in an immune cell may behave differently in a heart cell. Broad tissue measurements can blur those differences.
That makes a disease-associated DNA signal difficult to interpret. Finding an association is only part of the job. Researchers also want to know which variant is responsible, which gene it affects, whether it increases or decreases activity, and where that happens.
AlphaGenome 2 is designed to make those questions more specific. DeepMind describes it as a conditional model: a million-letter DNA sequence and a target biological condition go in together. The condition is used throughout the model, rather than treating every tissue output as a parallel prediction from the same sequence[1].
The distinction matters when reading a headline. A more detailed prediction is still a prediction. It can help a scientist choose an experiment without establishing that the predicted mechanism is the cause of a patient's symptoms.
What changed from the first AlphaGenome
The first model predicted many molecular readouts from a sequence across broad tissues and cell lines. DeepMind says that approach spread capacity across many outputs and made subtle differences between closely related cells harder to capture.
AlphaGenome 2 predicts for one condition at a time. Researchers can choose a condition from a reference catalog or derive a cell-state representation from their own single-cell measurements. This lets the question reflect the cell state under study, including states that are absent from the reference catalog[1].
DeepMind reports training on more than 107 million individual cell profiles alongside bulk data from human and mouse samples. Its launch summary describes coverage across 68 tissues. The point of that scale is a richer view of cellular differences, not a guarantee that every relevant disease state is adequately represented.
Nature's reporting explains the same change through the practical problem researchers faced with the earlier model: predictions that lacked precise cell-type specificity were less useful for some biological questions. That outside reporting supports the importance of the problem; it is not an independent replication of every model score[2].
What the reported improvement actually measures
DeepMind's clearest numerical example uses TenK10K, a map of genetic effects across 28 primary immune-cell types. It reports identifying 30% of causal variants, compared with 16% for the prior model fine-tuned on single-cell data. The corresponding figures for directional gene effects are 43% and 23%[1].
These numbers should stay attached to the dataset and task. They do not mean the model explains 30% of all genetic disease, nor that it is 43% accurate across every possible biological prediction. They describe selected evaluation outcomes against a specific comparator.
Likewise, “nearly twice as many” is a relative statement. It does not mean that a model has solved the entire problem because one evaluation improves substantially. The remaining missed cases, the confidence threshold and the mix of cells all matter when deciding whether a result is useful.
For a general example of why scores need their settings beside them, the November coding-model comparison separates benchmark evidence from an editorial recommendation. Genetics requires even more care because a better model score is several steps away from a clinical outcome.
A known IBD finding is a useful demonstration
The launch describes a variant linked to inflammatory bowel disease that increases activity of the ETS2 gene in gut macrophages. AlphaGenome 2 computationally reproduced the cell-specific pattern: the variant affects that gene in those immune cells rather than producing the same effect everywhere[1].
Nature reports that the biological finding had previously required substantial experimental work. Its article quotes a researcher involved in that earlier work describing the computational result as reassuring[2].
The right interpretation is a demonstration against an existing biological result. It is not evidence that the model independently discovered a new cure, eliminated the original experimental work or validated every similar prediction.
Reproducing a known mechanism is valuable because it gives researchers a concrete example to inspect. But it is still important to distinguish a retrospective illustration from prospective discovery on a new, unresolved disease mechanism.
New cell states can be supplied without retraining
A second important change is the ability to condition on new single-cell data. DeepMind says a small single-cell RNA-seq sample can be compressed into a cell-state representation and used to make predictions without retraining the full model[1].
The launch gives a vaccination-study example. With previously unseen data, the model distinguished activated immune T cells from resting cells and predicted a different effect for a variant associated with severe COVID-19 in the activated state.
This is a research capability, not a promise that any tiny sample provides a reliable answer. The input measurements still have to reflect the intended biological state. A poorly matched or low-quality sample can make the question itself less representative.
Before relying on a result, record what the conditioning data represents, how it was processed and whether the sample is appropriate for the comparison. “No retraining” changes the workflow; it does not remove data-quality requirements.
Imputation is not an observed measurement
AlphaGenome 2 can also predict missing experimental readouts. DeepMind describes using one kind of measurement, such as gene activity, to estimate another, such as chromatin accessibility. The model also adds DNA methylation to the readouts discussed in the launch[1].
This can help scientists decide where to spend limited experimental time. It can also be misread if predicted values are mixed with measured values without labels.
Keep those categories separate in a notebook, chart or dataset. An imputed value can be useful evidence for a hypothesis, but it should not acquire the status of an experiment merely because it appears in the same table.
A proposed research report should therefore identify the original measurements, the model inputs, the predicted outputs and the independent checks. That separation makes it possible to revise the conclusion when new experimental evidence arrives.
The drug-discovery question remains open
Better predictions can help connect a disease-associated region to a gene and a relevant cell type. But that is not the only evidence used to choose a drug target.
In a response to the launch, Genomics CEO Peter Donnelly says AlphaGenome 2 makes confident predictions for only a small subset of variants that can be queried. He also says its added value beyond other evidence, including proximity, colocalisation and rare-variant information, remains an open question. Genomics plans a detailed evaluation across its own GWAS data[3].
That is a useful limit to retain. The commentary comes from an interested industry practitioner, not a completed independent benchmark of the new model. Its value here is the practical question it raises: does the prediction add information beyond what a research team already knows?
A model can improve on its predecessor while still needing evaluation against an established scientific workflow. Those are different comparisons, and both can matter.
Availability has separate research and commercial routes
As of the November 11 announcement, DeepMind says AlphaGenome 2 is available for non-commercial research through the AlphaGenome API, with interactive Colab tutorials for expression conditioning. Commercial research is in private preview[1].
Do not turn those statements into a claim of unrestricted commercial access, a particular price or a clinical approval. The launch does not establish those things. Check the current access terms for the exact use before starting a project.
The announcement also says AlphaGenome 2 Skills in Google Antigravity are coming soon. “Soon” is not the same as an already available integration. A release summary should keep current access separate from a future tool announcement.
If a project involves sensitive genetic data, use the institution's approved data-handling process. A public model announcement is not permission to upload a person's data to an external service.
How to read a result before acting on it
A useful first check is to write the question in one sentence: which variant, in which cell state, affecting which readout? If that sentence is vague, a precise-looking model output may not settle the real research question.
Then preserve the input, model version, condition and confidence information. Compare the output with other genetic and biological evidence. Be explicit when evidence disagrees, rather than presenting every signal as support for a preferred story.
Finally, choose the next experiment or review step that could change the conclusion. This is proposed research practice, not a validation protocol tested in this article. The aim is to make a prediction useful without treating it as a result it has not earned.
The same caution applies to summaries of papers. A transcript review before summarization checks whether the source supports the summary. Here, the equivalent is checking whether a molecular prediction supports the stronger biological or clinical claim being made.
A more specific research tool, not a diagnostic shortcut
AlphaGenome 2's main contribution is its cell-specific framing. It lets researchers ask where a DNA change acts, instead of accepting a broad average that may hide the relevant biology.
The reported improvements and known-mechanism examples make that a useful research development to follow. The open questions are also concrete: coverage, confidence, added value beyond existing evidence and performance in new research settings.
For now, treat the output as a way to prioritize and interpret research. Preserve the cell condition, the evaluation setting and the uncertainty. None of those details should disappear when the result becomes a headline.
FAQ
What is AlphaGenome 2?
It is Google DeepMind's DNA sequence model for predicting molecular effects in specific cell types and states. It is intended for research, not clinical diagnosis.
How is it different from the original AlphaGenome?
It conditions predictions on a selected biological state rather than producing the same broad set of tissue outputs in parallel. Researchers can also supply new single-cell data to define a condition.
Does the improvement mean it can explain every disease variant?
No. Reported results describe particular tasks and datasets. They do not establish complete coverage of genetic variants or clinical usefulness for an individual.
Can it use a new cell state without retraining?
DeepMind says new single-cell gene-activity data can define a condition without retraining the model. The quality and relevance of that data still need checking.
Is AlphaGenome 2 available for commercial research?
The launch describes non-commercial research access through the API and a private preview for commercial research. Check current terms for a specific project.
Can AlphaGenome 2 diagnose a condition?
No. DeepMind says the models have not been validated or approved for clinical or diagnostic purposes.
