Markdown edition

# AlphaGenome 2 adds cell context to DNA predictions

> A research model predicts molecular effects in selected cells.

By BIGCHANGE.AI Editorial

Published: 2026-11-13T05:54:37.934Z
Updated: 2026-11-13T05:54:37.934Z
Canonical: https://bigchange.ai/blog/alphagenome-2-cell-specific-predictions

![Illustration of a researcher preparing a cell-specific DNA prediction.](https://bigchange.ai/api/media/file/genome-v3.png)
AI-generated illustration for BIG CHANGE.

DeepMind introduced [AlphaGenome 2](https://deepmind.google/blog/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease/) on November 11. Its API is available for noncommercial research, with commercial research in private preview. It is not validated or approved for clinical use.

## The big change

- AlphaGenome 2 adds a cell type or state to its DNA input. Unlike its predecessor’s fixed parallel outputs, biological context now shapes each prediction.
- The API tutorial compares an unchanged sequence with a variant in a selected blood-cell condition. This gives researchers a concrete workflow for inspecting predicted differences in gene activity.
- Earlier human-cell experiments illustrate why prediction and validation are separate. Researchers can alter a regulatory region in living cells and measure the response, rather than infer it from sequence alone.

## From sequence to a research question

The [operational guide](https://www.alphagenomedocs.com/getting_started.html) requires an API key and software package. Each call uses one condition; comparisons require separate conditions in repeated calls or batch items. Custom expression inputs can contain up to 256 human cells. This is an input limit, not a guarantee that a small sample represents a disease or population.

Its worked example selects CD14-positive blood monocytes, an immune-cell population, and a chromosome 2 variant that replaces C with T. The code centers a million-letter DNA interval on that change, predicts reference and alternate gene-activity tracks, then scores the effect and selects the GLB1L gene. Looking at the two tracks shows where the predictions differ; the score summarizes an estimated effect. These are tutorial instructions, not a model run performed for this article.

Before uploading biological data, a research team should establish its permission to use and disclose those data. Technical acceptance by an API does not itself establish consent or permission for a particular project.

## What the test establishes

The team’s [paper](https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease/alphagenome-2.pdf) evaluates 6,346 statistically selected variant–gene associations and 6,346 matched controls across 28 blood immune cell types. At 90% precision, recall was 30%, versus 16% for its single-cell-tuned predecessor. These are statistical labels, not individual experimental confirmations.

Precision measures how many selected predictions match positive labels; recall measures how many positives are found. A high precision can coexist with missed positives. Neither statistic, by itself, establishes performance in a different population or clinical setting.

The authors report underrepresented disease and transient states and underestimated effects. Their results do not establish clinical accuracy or make an unflagged variant harmless.

## Why experiments still matter

A [2024 Nature study](https://www.nature.com/articles/s41586-024-07501-1) offers a concrete experimental comparison. It investigated a regulatory region associated with inflammatory diseases and linked it to the ETS2 gene, about 290,000 DNA letters away. Such a region can influence a gene without being part of its protein-coding sequence.

The researchers used CRISPR–Cas9 to remove the regulatory region from primary human monocytes, then exposed the cells to inflammatory signals. They also disrupted ETS2 itself to investigate its role in macrophage inflammation. Changing the biological system and measuring its response supplies a different kind of evidence from a prediction. Those earlier experiments do not validate every new model prediction or establish a treatment for patients.

## Sources

- [DeepMind announcement](https://deepmind.google/blog/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease/) — Research access and model design.
- [Research paper](https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease/alphagenome-2.pdf) — Developer evaluation and limits.
- [API guide](https://www.alphagenomedocs.com/getting_started.html) — Operational example and input requirements.
- [Earlier ETS2 experiment](https://www.nature.com/articles/s41586-024-07501-1) — Indexed original methods; direct access unavailable.