JOURNAL / 2026.11.14
AlphaGenome 2 predicts DNA variant effects for specific cell states
DeepMind brings the context of specific cells into genetic effect prediction. Using researchers’ own data expands the questions they can investigate, though the quality of the answers depends on which signals have experimental support.
A DNA variant may deserve attention in one cell and go unnoticed in another. Choosing where to look for its effect is part of the scientific problem. AlphaGenome 2, announced by DeepMind on November 11, makes that context a model input. It is available through an API for non-commercial research; commercial access is in private preview.
In September, we examined AlphaGenome Atlas: the usefulness of having precomputed predictions at genome scale. This update changes another part of the work. The migration guide explains that AlphaGenome 1 returns signals from a catalog of tissues and assays, whereas the new version takes a sequence of up to a million bases and a biological condition. Comparing several conditions requires separate predictions. When designing an analysis, that means deciding which contexts to include and how to justify their selection.
The most promising feature is that this context can come from the researcher’s experiment. The official tutorial allows researchers to supply gene expression counts from their own human cells without fine-tuning the model. It requires raw counts and appropriate gene identifiers: a normalized matrix is not interchangeable with the expected input. In my view, the opportunity lies in connecting a local observation to genetic hypotheses that previously required settling for a cell category in the catalog. Data preparation becomes part of the validity of the question.
There is evidence of improvement, within a specific scope. In the technical manuscript, Figure 2e, AlphaGenome 2 recovers 30% of variants considered causal for expression in TenK10K, versus 16% for AlphaGenome 1 fine-tuned on single-cell data, at 90% precision. The comparison uses statistically mapped variants and selected controls; it does not measure diagnoses. The manuscript acknowledges underrepresented states and relationships learned from observations, without establishing that it predicts causal responses to perturbations. Model weights and code are promised upon final publication.
The kind of answer also matters. The imputation documentation distinguishes signals measured during training, inferences supported by informative modalities, and exploratory predictions without sufficient experimental constraints. Even the first category remains a prediction when querying a new sequence. For example, the guide warns that a transcription factor binding signal can look strong even when that factor is not expressed in the cell. A convincing curve can describe a biologically implausible hypothesis.
I would preserve that provenance alongside every result, including exports. A tool that gave all curves the same appearance would lose information needed to decide the next experiment. I would prefer it to show why a prediction deserves confidence, which measurement is missing, and what observation could refute it. Scientific interface design has consequences for the team’s reasoning here.
Imagine a group studying two states of an immune cell. It could compare the same variant in both, reserve some independent measurements before consulting the model, and then check whether the predicted difference appears. I would also include candidates with low scores: examining only the favorites would prevent us from learning what the filter misses. This is my proposal, not an experiment we have performed. The useful outcome would be a better decision about where to spend experimental effort, including cases where the model does not help.
The API repository notes that live predictions suit small analyses and that query rates depend on demand. Its terms exclude clinical decision-making. I would therefore begin with a bounded comparison, recording the version, cell context, and selection criteria. My view is that AlphaGenome 2 deserves attention for enabling more specific questions. Its value will grow if that specificity produces hypotheses other laboratories can challenge, rather than merely plausible explanations of known results.
Primary sources consulted on November 14, 2026. We have not run the model or reproduced its evaluations. The proposed uses and assessment are our own analysis.