Google DeepMind announced AlphaGenome 2 on November 11, 2026. The model predicts how DNA variants affect particular cell types. For researchers, the practical question is how to use that specificity to design a better experiment, while keeping prediction separate from biological evidence.
Start with a biological question
A useful analysis needs more than a list of variants. It needs a question about the biological context in which those variants might matter. For example, a research team could ask whether a candidate variant has different predicted effects in two relevant cell populations. That comparison would generate a hypothesis to test; it would not establish that the variant causes disease.
This changes how a project should be scoped. Instead of immediately screening every available condition, define the target gene or region, the candidate variants, and the cell populations that would make the result informative. Decide in advance what evidence would support the hypothesis and what result would weaken it. This makes a small analysis more useful than a large, poorly specified one.
What changes at the API level
AlphaGenome 2 makes each prediction for one biological condition. Researchers can select a reference condition or condition predictions on their own single-cell RNA-sequencing data. Comparing several conditions therefore needs explicit planning in the analysis.
The API separates predicting reference and alternative allele tracks from scoring variant effects. These outputs answer related but different questions: a score helps prioritize candidates, while inspecting the prediction tracks helps explain what changed and where.
A sensible first project
The following is a suggested evaluation workflow, not a report of an experiment performed for this article.
- Choose a narrow question and a manageable candidate set. Include known examples and appropriate controls rather than selecting only variants expected to produce a striking result.
- Check the reference genome, coordinates, alleles, and gene identifiers. Record the input files and software versions so that another researcher can reconstruct the analysis.
- Run a small pilot across the chosen conditions. Save both the numerical output and the information needed to interpret it.
- Inspect disagreements and negative results. A high score is a reason to investigate, not a substitute for an explanation. A weak prediction also needs interpretation in light of the selected condition and available evidence.
- Select follow-up experiments using the predictions alongside independent evidence. Define the validation criteria before expanding the analysis.
Keep measured and imputed evidence distinct
The documentation distinguishes tracks associated with observed training data from supported and exploratory imputation. That distinction should remain visible when results are filtered, plotted, or shared.
In a research report, place the prediction next to its context and limitations. Avoid giving visually similar plots the appearance of equivalent experimental support. Also separate three claims: the model predicts a molecular change; an experiment confirms that change; and the change explains a disease mechanism. Each requires its own evidence.
Access, capacity, and the November deadline
Non-commercial research access is available through the AlphaGenome API. Commercial research is in private preview, according to the launch announcement. AlphaGenome 2 is intended for research and has not been validated or approved for clinical or diagnostic use.
The documentation says non-commercial API use is free, with query rates varying according to demand. Live predictions suit analyses on the scale of thousands rather than millions; precomputed Atlas lookups allow higher rates. Free access should therefore not be treated as unlimited compute.
Google DeepMind is also inviting community use cases by November 30, 2026. Selected projects may qualify for additional quota and credits. The announcement does not specify an award amount or deadline time zone. Teams considering a submission should use the form linked from the announcement, check its latest terms, and avoid assuming an end-of-day cutoff in their own location.
What to look for next
For a research team, the most useful early result would be a reproducible analysis that identifies a worthwhile follow-up experiment. Before scaling up, ask whether cell-specific predictions actually change which candidates you would investigate, whether the result survives reasonable analytical choices, and whether the proposed validation can distinguish competing explanations.
That is a practical way to assess AlphaGenome 2: by the quality of the research decisions it supports, with experimental validation still doing the work of establishing the biology.
Information checked: November 12, 2026. This article reviews public documentation and proposes an evaluation workflow; it does not report independent model testing.