Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseAPI / Product · ActionableHabr AI→ original

AlphaGenome 2 predicts gene activity in a specific cell from a DNA stretch of up to 1 million letters

A single DNA mutation can strongly change a gene's activity in some cells and barely affect others. Google DeepMind's AlphaGenome 2 takes up to 1 million nucleotides (DNA letters) and a cell profile, such as a lymphocyte during inflammation. The first version knew only a fixed set of tissues. For biotech teams, it offers a fast path from a list of mutations to a hypothesis for the lab.

AI-processed from Habr AI; edited by Hamidun News
AlphaGenome 2 predicts gene activity in a specific cell from a DNA stretch of up to 1 million letters
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Key metric #1In the inflammatory bowel disease example, a variant boosted the ETS2 gene in macrophages
Key metric #2The Atlas of 9 billion mutations remains the tool for coarse filtering, followed by AlphaGenome 2
Key metric #3API access is available only for non-commercial research

The same single-letter change in DNA can strongly alter a gene's activity in some cells and barely affect others. It is like a line in a config file: harmless in a test environment, but it breaks the service in production because the environment is different. On November 11, Google DeepMind introduced AlphaGenome 2.

The model takes a DNA stretch of up to 1 million nucleotides (letters of the genetic code) and a description of a specific cell, such as a lymphocyte during inflammation. It then predicts how a mutation will change gene activity in that particular cell. If you build AI tools for biotech or follow the science, there is a practical technique to note here: the model is given context instead of an averaged picture across all cases.

Why cell context matters so much

Nearly all cells in the body contain the same DNA, but a neuron and a smooth muscle cell need different proteins. The difference lies in which genes are read, which is known as expression. Only 1–2% of DNA codes for proteins. The remaining 98% or so was once called "junk," but we now know it holds landing sites for regulatory proteins, genes for auxiliary RNAs, and elements that set the three-dimensional folding of chromatin (the complex of DNA and histone proteins). The expression profile shifts with immune signals, hormones, a cell's developmental stage, temperature, and even mechanical stress.

That is why a mutation outside the genes themselves can also cause disease: it changes a gene's activity, and the cell lacks a needed protein. Predicting such effects is still hard.

What changed from the first version

The first AlphaGenome came out last year. It accepted up to 1 million nucleotides and produced averaged profiles for a fixed set of tissues and cell lines from the training set. Think of a national weather forecast: you get the overall picture, but the model can say nothing about a specific backyard.

The second version adds a cell encoder (Cell encoder) on the input side. It converts an expression profile into a vector, a numerical description of the cell's state. The network takes this context into account in all its key layers. Training used data from the CZ CELLxGENE and Human Cell Atlas projects, which describe millions of individual cells. You can pick a profile from a ready-made database or upload your own. As a result, the prediction is built for both the cell type and its state.

The architecture stays the same: a U-Net, where convolutional layers (CNN) find short local patterns and transformers link regions up to a million bases apart. The model outputs an expression profile and a map of regulatory changes at single-nucleotide resolution. The map shows which regions will remain open for reading, where transcription factors can bind, and how chromatin contacts in the nucleus will change.

An example and where it fits in the workflow

DeepMind's announcement includes an example involving inflammatory bowel disease. The model reproduced the result of an earlier experimental study and singled out a variant in a distant regulatory region. This mutation selectively boosts the activity of the ETS2 gene in monocytes and macrophages and barely affects other cells. For a researcher, this is a short path from a list of suspect variants to a testable hypothesis: which gene is affected, in which cells, and what to check in an experiment.

AlphaGenome 2 does not replace AlphaGenome Atlas, the database launched in September. In it, the model precomputed the effects of 9 billion single-letter mutations: the genome has about 3 billion bases, and each has three possible substitutions. Running that many variants through a heavy context-aware network for every cell state is too resource-intensive. So researchers first filter candidates coarsely using the Atlas, then run the most interesting ones through AlphaGenome 2. It can also take more complex combinations of mutations.

The model is part of DeepMind's biology ecosystem: AlphaMissense assesses the danger of amino acid substitutions, AlphaFold predicts protein structures, and AlphaProteo designs proteins for a chosen target.

Limitations and practical takeaways

Even accurate predictions require experimental confirmation in clinically significant cases. The ETS2 example shows that the model repeated an already known result, so it cannot be counted as an independent discovery. You can try AlphaGenome 2 through the API, but only for non-commercial research. Only the first version has its code and weights openly available.

For teams building tools for labs, the takeaway is this. The pattern of a cheap reference lookup as the first step and an expensive context-aware model as the second applies beyond genomics. If you are building a product for target discovery or variant analysis, it makes sense to study the API now and plan how to fit it into your pipeline. DeepMind has not yet announced commercial terms, so it is too early to build a product solely on AlphaGenome 2.

Story

Key Questions Answered

Why cell context matters so much?

Nearly all cells in the body contain the same DNA, but a neuron and a smooth muscle cell need different proteins. The difference lies in which genes are read, which is known as expression. Only 1–2% of DNA codes for proteins. The remaining 98% or so was once called "junk," but we now know it holds landing sites for regulatory proteins, genes for auxiliary RNAs, and elements that set the three-dimensional folding of chromatin (the complex of DNA and histone proteins).

What changed from the first version?

The first AlphaGenome came out last year. It accepted up to 1 million nucleotides and produced averaged profiles for a fixed set of tissues and cell lines from the training set. Think of a national weather forecast: you get the overall picture, but the model can say nothing about a specific backyard. The second version adds a cell encoder (Cell encoder) on the input side. It converts an expression profile into a vector, a numerical description of the cell's state.

What are the key details on An example and where it fits in the workflow?

DeepMind's announcement includes an example involving inflammatory bowel disease. The model reproduced the result of an earlier experimental study and singled out a variant in a distant regulatory region. This mutation selectively boosts the activity of the ETS2 gene in monocytes and macrophages and barely affects other cells. For a researcher, this is a short path from a list of suspect variants to a testable hypothesis: which gene is affected, in which cells, and what to check in an experiment.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…