AlphaGenome turns DNA variants into testable research hypotheses

Predicted molecular effects can help prioritize experiments, provided researchers keep clinical interpretation separate.

Two DNA sequences can differ by one letter while leaving researchers with a much larger question: what, if anything, changes in the biological system? Sequence-to-function models offer one way to organize that uncertainty. They predict molecular measurements associated with a sequence and compare the predictions before and after a variant. The result can help decide which experiments deserve scarce laboratory time.

AlphaGenome was introduced in June 2025 and its research was published in Nature in January 2026. Google DeepMind's research implementation describes predictions across gene expression, splicing and chromatin-related modalities, with long sequence context. The repository supplies model and analysis code. These are resources for studying regulatory effects, not an established clinical interpretation of an individual person's genome. [Research publication](https://www.nature.com/articles/s41586-025-10014-0), [implementation](https://github.com/google-deepmind/alphagenome_research/blob/main/README.md), [announcement](https://deepmind.google/blog/alphagenome-ai-for-better-understanding-the-genome/).

The distinction between a molecular effect and a health outcome is crucial. A model may predict that a sequence alteration changes a particular signal in a particular biological context. It does not follow automatically that the alteration causes a disease, determines its severity or identifies a useful treatment. Each additional link needs its own evidence. Collapsing them into a single “risk” score hides the unanswered questions.

A practical research project should start with a narrowly defined hypothesis. For an illustrative example, a laboratory might ask whether a set of variants near a gene merits follow-up in an expression assay. Choose the relevant predicted output before viewing the results. Searching every available track and reporting only a striking difference makes it harder to distinguish a useful lead from a selection artifact.

Prepare the input carefully. Record the reference genome version, coordinates, sequence context and variant representation. Confirm that the reference sequence matches the variant record. A coordinate mismatch can generate an apparently sophisticated analysis of the wrong sequence. Keep a reproducible input manifest and model version so another researcher can repeat the comparison.

Generate reference and altered-sequence predictions under the same conditions. Examine the surrounding pattern, rather than only the largest numerical difference. Compare the result with known controls and with a simpler baseline where appropriate. A useful research report explains what was compared, which output was selected and why that output corresponds to the original question.

The next stage is independent experimental work. Select candidates with a protocol that includes negative controls and a realistic way to observe the proposed effect. If the experiment does not confirm the prediction, preserve that result. A collection containing only successful confirmations teaches readers very little about the model's practical reliability or the cost of following its suggestions.

Generalization needs deliberate attention. A model's useful behavior in one cell context or measurement modality does not establish equivalent reliability elsewhere. Ask whether the intended experiment resembles the conditions represented in development and evaluation. Treat unfamiliar sequence patterns and poorly represented contexts as reasons to gather more evidence, not merely as invitations to choose a higher threshold.

Access should be evaluated as part of reproducibility. Distinguish source-code terms, model-weight terms, hosted API conditions and any restrictions on commercial work. An accessible repository does not settle all four. Local compute, data preparation and laboratory confirmation can dominate cost even when some research access is free; the sources here do not verify an all-in price.

Genomic privacy also applies to research tools. Public examples and authorized reference sequences are different from identifiable participant data. Determine which inputs may be sent to a hosted system under the study's consent and governance arrangements. Store the minimal information needed to reproduce the analysis and restrict access to sensitive records.

Prioritization also needs a stopping rule. Before ordering assays, decide how many candidates the available budget can test and how controls will be selected. Keep that rule unchanged after seeing predictions. This makes negative results informative instead of allowing researchers to keep searching until one attractive confirmation appears.

The strongest use of AlphaGenome is a disciplined narrowing of experimental uncertainty. It can give scientists a structured candidate hypothesis and a reproducible comparison to investigate. The scientific gain comes when that prediction meets independent measurements, including disappointing ones, and the record makes clear which conclusions were observed and which remain proposed.