Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Everyday tasks

AlphaGenome 2 reads variants per cell type

Written November 12, 2026

Every cell in a body carries the same DNA, but a variant that matters in a gut immune cell may do nothing in a heart cell. The first AlphaGenome read variants against broad tissues. AlphaGenome 2, announced November 11, 2026, reads them against the specific cell type — and that change roughly doubles how many expression-changing variants it finds, on DeepMind's own figures.

The redesign: one condition at a time

DeepMind's announcement explains the architectural change plainly. The first model took a million-letter DNA sequence and predicted results across thousands of experiments from many tissues in parallel — efficient, but its capacity was spread thin, and subtle differences between closely related cell types were hard to capture. AlphaGenome 2 is a conditional model instead: alongside the sequence, it takes one biological condition — a tissue, cell type or state — as an input at every stage, and predicts for that condition. It was trained on bulk and single-cell data from human and mouse samples, including more than 107 million individual cell profiles across 68 tissues. It predicts the same core readouts as before — gene expression, histone modifications, chromatin accessibility — and adds DNA methylation, a readout DeepMind says was widely requested.

What the numbers say

On the TenK10K dataset, a single-cell map of 28 primary immune cell types, DeepMind reports that AlphaGenome 2 identified nearly twice as many causal variants as the original model fine-tuned on single-cell data and other tools: 30% against 16% for causal variants, and 43% against 23% for the direction of the gene effect. It also tracks how expression changes continuously as cells mature, for example as human blood stem cells become red blood cells and immune lineages. As a worked example, DeepMind revisits a published finding on inflammatory bowel disease, where a variant in a distant regulatory switch overactivates the ETS2 gene in gut macrophages: run across gut immune cell types, the model singled out that variant and showed the boost landing in macrophages, with other cell types and genes largely unaffected.

New cell states, no retraining

The conditional design buys a second trick. A scientist can hand the model a gene-activity snapshot from a few dozen cells of a type it has never seen; the snapshot is compressed into a cell state, and the model predicts for that state without retraining. DeepMind's example comes from a COVID-19 mRNA vaccine study: the model separated vaccine-activated immune T cells from resting ones, and showed a variant linked to severe COVID-19 disrupting gene activity more strongly in the activated state — an effect invisible in resting cells. DeepMind has also precomputed scores for trait-associated variants across hundreds of human cell and tissue conditions, covering traits from height to type 2 diabetes, and extended enhancer-to-gene mapping into primary human cell types such as brain astrocytes.

How to reach it

AlphaGenome 2 and interactive Colab tutorials are available for non-commercial research through the AlphaGenome API, and in private preview for commercial research. DeepMind also says AlphaGenome 2 Skills are coming to Google Antigravity soon.

Limits

Related playbooks

Sources

Back to: Everyday tasks playbooks hub · Also: AlphaProof Code verified Signal in Lean · T21: a research assistant that reads results first