top of page

Google AlphaGenome 2 targets disease at the cellular level

44 minutes ago
11 min read
Illustrative photograph. Photo: Zifeng Xiong / Pexels.

Google DeepMind announced Google AlphaGenome 2 on November 11, shifting DNA predictions from broad tissue patterns toward specific cell types and states. The research model addresses a consequential gap: identifying a genetic change matters less when scientists cannot establish where its effects appear. More detailed predictions promise better experiments, but they do not establish that a treatment will work.

The central comparison is with the original AlphaGenome, which predicted molecular activity across many predefined outputs. Its successor makes biological context an explicit input. That changes the research question from whether a DNA change affects activity toward which cells experience that effect.

For genetics researchers and drug discovery teams, the distinction puts pressure on how computational evidence gets evaluated. A convincing prediction must connect a genetic change, a molecular consequence, and a relevant biological setting. The announcement strengthens that approach while leaving experimental confirmation essential.

Google AlphaGenome 2 brings cellular context into the prediction

According to DeepMind’s cell-specific model announcement, the system processes a million-letter DNA sequence alongside a chosen biological condition. That condition can describe a tissue, cell type, or cell state. DeepMind reports training on more than 107 million individual cell profiles across 68 tissues, using human and mouse data.

The important change is that cellular context becomes part of the question the model receives.

A genetic variant, a difference in a DNA sequence, does not necessarily produce the same consequence everywhere. Gene expression, the activity involved in using a gene’s instructions, provides a measurable consequence researchers can investigate. A useful prediction therefore needs a setting as well as a sequence.

Consider what a research team actually needs from an answer. A broad indication that a region matters can help identify a starting point. It leaves substantial work before anyone can design an informative experiment. Researchers still need to decide which cells to examine and which outcome to measure.

Adding context makes that decision more explicit. The same sequence can be evaluated under different biological conditions, producing hypotheses that researchers can compare. A proposed mechanism becomes more specific when it identifies where an effect should appear and where it should remain absent.

That specificity also creates a more demanding standard of evidence. If a prediction names a particular cell type, experiments can test that claim directly. A result that appears everywhere would tell a different story from an effect concentrated in the predicted setting.

This is why the release deserves attention beyond its version number. The design directs attention toward biological differences that broad summaries can obscure. It also makes those differences available for scrutiny rather than leaving researchers to infer them from a general score.

There is an important boundary to maintain. Providing a cellular condition does not mean the model recreates an entire living cell. Researchers should interpret its outputs within the tasks evaluated, rather than treating them as a complete account of disease.

For readers following AI research, that boundary helps distinguish a specialized scientific model from a general assistant. Its value rests on whether its predictions support useful scientific decisions. A fluent explanation or attractive visualization cannot substitute for that evidence.

The immediate question is therefore practical: does the added context help researchers choose experiments that reveal something they would otherwise miss? That is a more meaningful standard than whether the model produces a richer-looking output.

Disease research needs a location for genetic effects

The researchers under pressure are those turning genetic associations into testable explanations of disease.

A statistical association between a DNA region and a disease can identify an important lead. It does not automatically explain the biological chain connecting them. Researchers need to work out which change matters, what it affects, and why that consequence becomes relevant in particular cells.

Each unresolved link complicates the next decision. A team choosing an experimental system needs confidence that the system represents the proposed mechanism. A team considering a therapeutic target needs evidence that changing its activity addresses the disease process.

Cellular context makes those questions harder to avoid. If a predicted effect depends on a specific state, testing an unrelated condition might produce an uninformative result. Conversely, a favorable result in a convenient laboratory setting might say little about the setting that matters.

The pressure is methodological rather than a claim that laboratories must immediately adopt a new product. Research teams need to justify whether their chosen systems can answer their questions. A cell-specific prediction offers another hypothesis to assess when making that choice.

It also changes how a negative result should be interpreted. Failure to observe an effect might challenge the proposed mechanism. It might instead expose a mismatch between the experiment and the biological context being claimed. Distinguishing those possibilities requires careful experimental design.

The potential benefit is a better sequence of decisions. Researchers could begin with a broader set of genetic leads, identify plausible cellular settings, and select experiments that discriminate between competing explanations. That is a possible workflow improvement, not a measured outcome for every project.

For drug discovery teams, a detailed mechanism can inform more than target selection. It can shape questions about where an intervention needs to act and what evidence would justify further investment. Those remain development questions requiring evidence beyond a molecular prediction.

Specificity also makes disagreement more productive. Instead of arguing over whether a model considers a variant important, researchers can compare its proposed cellular setting with other evidence. An explicit hypothesis gives collaborators something concrete to challenge.

The same discipline applies to scientific knowledge management. Teams need to retain the distinction between predicted effects, observed effects, and interpretations carried forward between projects. A searchable knowledge base can help preserve that evidence trail.

The broader implication is straightforward. Better predictions should improve the questions researchers ask before an experiment. Their usefulness becomes clearer when those questions produce informative answers, including answers that contradict the original model.

The ETS2 example connects prediction with an existing discovery

One concrete example concerns ETS2, a gene implicated in inflammatory biology. A macrophage inflammation study, published in Nature in June 2024, investigated a disease-associated region outside protein-coding genes. The researchers connected that region to increased ETS2 activity in macrophages, immune cells involved in inflammatory responses.

DeepMind says its new model computationally reproduced the relevant variant’s selective effect in gut macrophages. That makes the example useful because it connects a prediction with a biological mechanism investigated through experiments.

Recovering an established mechanism is an encouraging test, but discovering an unknown mechanism demands a different evaluation.

The difference concerns what researchers know when they begin. In a retrospective test, an experimental finding provides an answer against which predictions can be compared. In a prospective test, researchers must choose what to investigate before knowing which explanation will survive.

Both forms of evaluation matter. Recovering a known finding can show that a model captures a relevant relationship. Prospective work asks whether that capability helps identify useful leads when the answer is unavailable.

Nature’s cellular genetics coverage reported James Lee’s response to the computational result. Lee, an immunologist and geneticist at the Francis Crick Institute who led the earlier study, described it as reassuring. His reaction recognizes agreement with experimental work without establishing that computation replaces it.

That distinction becomes especially important when comparing research speed. Generating a hypothesis and establishing a biological mechanism are different activities. An experiment can reveal unexpected effects, challenge assumptions, and expose complications that a successful retrospective prediction does not address.

The ETS2 case also illustrates why location matters. A prediction about a gene alone leaves the cellular setting unresolved. Connecting the proposed effect with a relevant immune population makes the hypothesis more informative and the follow-up experiment more focused.

It would still be a mistake to turn that result into a treatment claim. Evidence that a gene participates in disease does not establish that a particular intervention will be effective or tolerable. The original experimental paper itself discusses challenges around therapeutic approaches and effects in other tissues.

For evaluating AI, the useful lesson is about matching claims to evidence. A recovered molecular relationship supports a claim about that relationship. It does not automatically support conclusions about patient outcomes, development timelines, or the number of unsuccessful experiments avoided.

Researchers can build on the example by asking what would count as a harder test. Would the model prioritize the correct explanation among plausible alternatives? Would its preferred cellular setting hold up in new samples? Would following its ranking outperform another selection method?

Those questions preserve what is interesting about the result. They turn a striking example into an evaluation agenda, with room for both successful predictions and informative failures.

AlphaGenome 2 vs AlphaGenome is a question of context

The clearest AlphaGenome 2 vs AlphaGenome comparison concerns how the models handle biological conditions. The earlier system produced predefined outputs in parallel. The successor conditions its predictions on a selected setting, according to DeepMind.

That comparison belongs within an established research field. The Borzoi research paper, published in Nature Genetics in January 2025, describes predicting cell-type-specific and tissue-specific RNA sequencing coverage from DNA. RNA sequencing measures RNA molecules to characterize gene activity.

The competitive question is which approach supports the research task under evaluation, rather than which model has the broadest headline.

Borzoi provides a useful reference because it prevents cell-related prediction from sounding like an entirely new category. Researchers have already pursued links between sequence and molecular activity. The relevant comparison concerns the details of the task, inputs, outputs, and evidence.

A model can be useful for one experimental decision without being preferable for every project. Teams might need predictions for well-characterized settings, comparisons across conditions, or analysis of newly collected observations. Those requirements should determine how models get evaluated.

The same principle applies when comparing the original and updated AlphaGenome systems. A newer architecture creates a reason to test performance, not permission to assume universal superiority. Researchers need to identify which tasks improved and which limitations remain consequential.

DeepMind’s AlphaGenome 2 paper accompanies the release and provides the technical basis for assessing its claims. In the company’s TenK10K evaluation across 28 immune cell types, reported causal-variant recall rose from 16% to 30%. Reported recovery of directional gene effects rose from 23% to 43%.

Recall describes how many relevant examples an evaluation recovers. Precision describes how many selected predictions are correct. Those concepts need to remain separate when interpreting a headline improvement.

The reported figures support a claim about performance on a molecular prediction benchmark. They are not percentages of patients diagnosed correctly. They also should not be read as the probability that any particular proposed treatment will succeed.

The AlphaGenome 2 vs AlphaGenome comparison is therefore strongest when kept close to the evaluated task. It becomes weaker when expanded into claims about all diseases, all cellular settings, or every competing method.

For research buyers, the appropriate response is a comparison matched to their own work. They should ask whether an improvement changes an experimental decision, whether that decision can be tested, and whether the result remains reproducible.

That approach leaves room for complementary tools. Agreement across methods can help identify a hypothesis worth investigating, while disagreement can expose assumptions that deserve attention. Neither outcome eliminates the need to assess the underlying evidence.

Better recall still leaves substantial uncertainty

An improved benchmark is evidence of progress on its task, not a complete measure of biological understanding.

The reported recall figures immediately raise a practical question: what happens to relevant effects the model does not recover? A researcher needs to know whether an unflagged variant should receive lower priority or remain unresolved.

That distinction matters because an absence of prediction is easy to overinterpret. A model might fail to identify an effect for reasons unrelated to whether the effect exists. Researchers should avoid turning incomplete computational evidence into a definitive negative conclusion.

Positive predictions require similar care. A ranked candidate can look persuasive when accompanied by a detailed proposed mechanism. The detail makes the hypothesis easier to test, but it does not turn prediction into observation.

Another issue is the comparison setting. Evaluations can differ in the examples selected, the outcomes measured, and the rules used to classify a prediction as successful. A headline percentage does not carry all that context.

Independent researchers can help establish how well the reported gains travel beyond the original evaluation. The important question is not whether every new experiment matches the launch results. It is whether performance remains useful across clearly defined tasks and conditions.

This is also where a simple reading of AlphaGenome 2 explained can become misleading. Explaining the architecture helps readers understand the proposal. Assessing the evidence requires asking which claims were tested, what remained outside scope, and what alternative explanations survived.

A particularly useful evaluation would record predictions before running follow-up experiments. That would make it harder to select only favorable examples after the outcome becomes known. Researchers could then compare the model’s priorities with competing methods or existing practice.

Failures should remain visible in that process. An incorrect cellular assignment might reveal a limitation different from an incorrect direction of effect. Combining every outcome into a single success score could obscure what researchers most need to understand.

Commercial teams also need to separate scientific performance from operational suitability. A model can produce useful hypotheses while leaving questions about access, reproducibility, and permitted use unresolved for a particular organization.

DeepMind explicitly states that AlphaGenome and AlphaGenome 2 are research tools without validation or approval for clinical or diagnostic use. That limitation should remain attached to downstream descriptions of their capabilities.

The appropriate conclusion is neither dismissal nor a promise of immediate medical benefit. Researchers have a more specific computational proposal to examine. Its broader value depends on what happens when teams use it to make decisions and then test those decisions.

Watch independent tests, new mechanisms, and usable access

The next phase for Google AlphaGenome 2 should be judged through concrete evidence rather than increasingly expansive descriptions. Over the coming months, independent evaluation, prospective experiments, and usable access will provide the most informative signals.

The strongest evidence will connect a prediction to a decision and then to an experimentally assessed outcome.

First, watch for evaluations from groups outside the model’s developers. Useful studies should describe the task, comparison methods, and conditions under which predictions succeed or fail. A reproducible gain on an independent dataset would strengthen confidence that the reported improvements generalize.

A weaker result would not automatically make the model useless. It would narrow the settings in which researchers should trust its rankings. That information is valuable when teams need to decide where computational guidance deserves further experimental investment.

Second, watch for newly investigated mechanisms rather than additional reconstructions of established findings. A compelling report would identify the model’s hypothesis before testing and explain how it influenced experimental selection.

The result should include enough detail to distinguish a successful prediction from a favorable interpretation after the fact. It should also show whether the cellular context contributed something useful beyond a general indication that a DNA region mattered.

This signal would directly test the release’s central proposition. If context-specific predictions repeatedly help researchers choose informative experiments, the case for adopting them strengthens. If they rarely change decisions, improved benchmark scores would have a narrower practical meaning.

Third, watch whether access supports sustained research. DeepMind announced noncommercial availability through an application programming interface, a service that software can query, alongside a commercial private preview. Its research API documentation provides an entry point for assessing the available workflow.

Availability alone does not answer every adoption question. Researchers need to determine whether the service supports their intended analyses and whether they can preserve enough information to reproduce results later.

Commercial organizations should likewise distinguish a private preview from broad production availability. The relevant signal is whether access develops into a dependable research workflow under clearly understood terms.

These observations also offer a useful filter for future announcements. Independent performance tests assess generalization. Prospective experiments assess whether predictions improve scientific choices. Practical access determines whether researchers can carry those gains into ongoing work.

None requires treating model output as clinical evidence. Together, they would provide a clearer account of where the system helps and where experimental uncertainty remains.

Before treating Google AlphaGenome 2 as evidence for a disease mechanism, read the evaluation alongside the underlying biological research. Ask which cellular condition was tested, which outcome was measured, and which alternatives remain plausible. For a research team, the next useful step is a bounded comparison using questions it already understands well. Record predictions before inspecting experimental answers, retain unsuccessful cases, and separate molecular agreement from therapeutic interpretation. That approach gives the model a fair test while keeping the evidence traceable. The question to carry forward is concrete: does adding cellular context help researchers choose a better experiment, and does that experiment support the proposed mechanism?

Give every agent the context to do better work

Connect your agents to the knowledge, decisions, and history already organized in remio.

remio currently supports Windows 10+ (x64) and Macs with Apple silicon.

Your AI Partner at Work
Get more done with remio

Plan. Create. Deliver.
All in one place.

bottom of page