AlphaGenome 2 adds cell type and cell state to the question of what a DNA variant might do. The advance could help researchers prioritise experiments, but its predictions remain one source of evidence rather than a demonstration of disease causation or clinical usefulness
A change in DNA does not have to alter a protein directly to matter. It can affect the regulation of a gene, and its consequences may depend on the cell in which that regulation occurs. That makes a seemingly simple question—what does this variant do?—a question about biological context. On 11 November, Google DeepMind announced AlphaGenome 2, a model designed to predict molecular effects for specified cell types and states. The advance is worth examining because it changes the structure of the prediction task, bringing an explicit representation of cellular context into the model alongside the DNA sequence.
DeepMind says the model was trained on more than 100 million individual cell profiles across 68 tissues. Access is through the AlphaGenome API for non-commercial research, with commercial research in private preview. The developer explicitly states that the system has not been validated or approved for clinical or diagnostic purposes. Those boundaries are part of the release, not qualifications to postpone until after discussing its promise. AlphaGenome 2 is a research instrument for investigating molecular activity. Its potential importance lies in improving which hypotheses scientists examine and which experiments they prioritise, while leaving substantial work between a model output and a dependable biological conclusion.
Cellular context becomes an input
The architectural change is conceptually clear. According to DeepMind, the earlier model produced fixed parallel outputs across tissues and cell lines. AlphaGenome 2 instead takes a million-letter DNA sequence together with a biological condition, predicting for one condition at a time. The condition can come from a reference catalogue or a representation derived from single-cell measurements. That representation informs the network throughout its architecture. The important point is that the model is asked a more specific question. A prediction is attached to a cellular setting rather than presented as if a stretch of DNA had one uniform molecular effect throughout the body.
An everyday analogy is a set of instructions used by different departments in an organisation. The text can be the same while different roles make different parts relevant. The analogy is only illustrative: cells are not office departments, and gene regulation is not a document-management problem. It helps explain why averaging across contexts can conceal an effect confined to one setting. For a scientific model, representing context explicitly offers a way to investigate such differences. Whether it does so reliably remains an empirical question, requiring appropriate measurements and comparisons rather than confidence in the analogy itself.
From a genetic association to a testable mechanism
In his assessment of the release, geneticist Peter Donnelly describes the practical challenge for research: identifying the relevant gene, the direction of its effect and the cell types in which it acts. A disease-associated region does not automatically supply those answers. His commentary welcomes AlphaGenome 2 while emphasising that its additional value beyond existing genetic evidence still needs evaluation. He also notes that confident predictions cover only a subset of queryable variants. This is an expert assessment from a company planning to evaluate and integrate the model, rather than a completed independent replication.
A possible use is to bring a more focused hypothesis to an experiment. In an illustrative case, researchers might begin with a region associated with a trait, then ask whether a particular change is predicted to increase or decrease activity of a nearby gene in a relevant cell type. The result could help organise the next investigation. It would not establish that altering that gene will treat a disease, nor that the predicted cellular effect explains the association. Keeping those steps separate gives the model a useful role without assigning it a conclusion that requires several different kinds of evidence.
Predicting a difference is not observing one
Sequence-to-function models support an intuitive comparison: obtain a prediction for a reference sequence, obtain another for an alternative sequence and examine the difference. Donnelly outlines that approach in his commentary. The numerical difference is the model’s estimate of a molecular change, not a laboratory observation. That distinction matters even when a prediction is detailed and visually persuasive. A large collection of predicted values can look like an experimental atlas while retaining the assumptions and limitations of the system that generated it. The application should preserve the origin of each value so that a downstream analysis does not silently treat predictions as measurements.
There is also a distinction between a model’s ability to rank candidates and its ability to estimate an effect accurately. A method might help put promising cases nearer the top of a list while still being uncertain about their precise consequences. A useful evaluation should make clear which capability it measures. An illustrative prioritisation workflow could tolerate some false leads if subsequent experiments can resolve them efficiently, whereas another use might require a much higher level of confidence. The acceptable error depends on the purpose and the cost of acting on a mistaken prediction. A broad claim that a model is better needs to be translated into that setting. An informative comparison could also ask whether the model adds useful information to a simpler method already used by the research team. If both approaches prioritise the same candidates, improved molecular detail might still help explain them, but the practical benefit differs from finding candidates that would otherwise be missed. This proposed comparison separates explanatory richness from additional predictive value. Both can matter, yet they should be reported distinctly so that an attractive biological narrative does not stand in for evidence of a better research decision.
A benchmark result needs its denominator
A report that a model identifies more variants can sound like a general increase in biological understanding. The interpretation depends on which variants were tested, what was treated as ground truth and which confidence threshold was used. It also depends on the baseline: an earlier model evaluated under different conditions would answer a different comparison from one adapted to the same data. These are questions for reading the evidence, rather than allegations about the release. They help distinguish an improvement on a defined task from a claim that the same improvement applies across every disease, cell type or form of genetic variation.
Selective confidence is particularly important. A model that makes few high-confidence calls may provide valuable information in those cases, while leaving many other cases unresolved. Donnelly cautions that a lack of a confident call should not be interpreted as evidence of no effect. For a researcher, that means preserving an unknown category rather than forcing every variant into a useful-looking classification. Evaluation should therefore examine both reliability among the calls made and coverage across the cases of interest. Otherwise, a strong score on a narrow subset could be mistaken for comprehensive interpretation of a much larger search space.
New cell states make transfer a central question
DeepMind describes using single-cell expression measurements to condition predictions for states absent from training, without retraining the model. This is a developer-reported capability, not a guarantee of transfer to any new dataset. It raises an important technical question: how well does the representation of a new state capture the information the predictor needs? A model can accept an input in the expected format while still encountering a biological or measurement setting unlike the ones on which its performance was established. The distinction between interface compatibility and validated generalisation is as important here as in other scientific applications of AI.
A careful assessment could compare results across experimental sources, cell states and regions of the genome, looking for systematic changes in performance. It could also examine how uncertainty behaves when relevant measurements are sparse or noisy. These are suggested evaluation directions, not tests conducted for this article. The purpose would be to discover where the new conditional design extends reliable prediction and where further data are required. A model’s ability to generate an answer for a new condition is useful only when researchers can judge what evidential weight to give that answer. That judgement requires more than the fact that inference completed successfully.
Access is concrete; the scientific payoff remains to be tested
The AlphaGenome Python package’s v2.0.0 release provides a dated software reference alongside the announcement. That makes the release more concrete than a proposal without an access path, while an API client should not be confused with downloadable model weights. Reproducible research still requires recording the model or service version, requested condition, input sequence and analysis choices. If those details change, a later result may be a different experiment even when the research question sounds the same. Access and documentation enable scrutiny; they do not replace it.
The larger opportunity is to make the relationship between genetic evidence and cellular biology more precise. AlphaGenome 2 offers a design in which context is part of the prediction, potentially helping researchers ask better-targeted questions. Its value will depend on what survives comparison with measurements, existing methods and independent assessment. For now, the warranted conclusion is a technical one: a newly available research model attempts to resolve DNA effects at a more specific cellular level. Turning that capability into reliable scientific knowledge will require the same discipline the model is meant to support—clear hypotheses, appropriate comparisons and a willingness to leave uncertain cases unresolved.
Image credit: Scientist using a pipette in a laboratory by Julia Koblitz, used under the Unsplash License. Illustrative photograph unrelated to Google DeepMind, AlphaGenome 2 or the researchers discussed. Resized for this site; WordPress may crop display versions.
