AlphaGenome 2: what it is and how to get access

AlphaGenome 2 is Google DeepMind's new DNA sequence model. It predicts how a change in DNA affects one specific cell type or cell state, where the first version could only answer at the level of a whole tissue. It is available now through the AlphaGenome API for non-commercial research, with a private preview for commercial work.

@GoogleDeepMind announced it on November 11 with a simple framing: your heart, brain and immune system all carry the same DNA, so the interesting question is what each cell does with it. @demishassabis followed with "Super proud to announce AlphaGenome 2 - our new DNA sequence model!" and pointed academics to the API.

What changed from the first AlphaGenome

The original model read a stretch of DNA a million letters long and predicted a large batch of tissue-level readouts at once. According to DeepMind's announcement, version 2 is a conditional model. You give it the same million-letter sequence plus one condition, such as a tissue, a cell type or a cell state, and it predicts for that condition only.

The condition can come from DeepMind's reference catalog or from a lab's own single-cell data, passed through what the team calls a cell encoder. A follow-up post in the @GoogleDeepMind thread says the model "adapts to lab experiments on the fly with no retraining needed," using data from a few dozen cells. The examples given are immune cells responding to a vaccine and gut cells involved in inflammation.

DeepMind says it trained the model on more than 107 million individual cell profiles from human and mouse samples, across 68 tissues. It also adds DNA methylation as an output and can fill in one type of measurement from another.

The numbers DeepMind is reporting

@pushmeet, one of the four authors on the announcement, posted the headline results. They come from TenK10K, a dataset covering 28 primary immune cell types.

Read those carefully. They are DeepMind's own figures, and 30% means the model still misses most causal variants. It is nearly double the old score, and it is also a long way from a diagnostic tool. DeepMind says as much: both AlphaGenome models are for research use only and have not been validated or approved for clinical or diagnostic purposes.

The announcement also describes the model reproducing a known mechanism in inflammatory bowel disease involving the ETS2 gene. I read that as a check against published biology. Nothing new was found there.

How to get access

Here is what the announcement lists:

• Non-commercial researchers can use it today through the AlphaGenome API. Documentation and sign-up are at alphagenomedocs.com.
• Commercial research is in private preview. The blog post gives no price and no general release date.
• DeepMind mentions interactive Colab tutorials for expression conditioning.
• There is a forum at alphagenomecommunity.com and a call for use cases that closes November 30. Selected projects may get extra quota or credits.
• AlphaGenome 2 Skills for Google Antigravity are promised "soon." They are not out yet.

If you are not a geneticist, there is nothing here for you to click on. This is a tool for labs, and it has no chat window.

My read

This is the third DeepMind research release we have covered this month, after AlphaProof Code and the 1024-bit RSA result. While Google's consumer side ships models like Gemini 4 Hydrogen, the science group keeps putting out narrow systems that do one hard thing and say how often they get it wrong.

I think the access terms deserve a closer look than the benchmark. Academics get the API now. Drug companies, the people who would pay for it, wait in a private preview with no public price. To me that reads as Google seeding the research community first and settling the commercial terms later.

The Antigravity detail is the other thing I would watch. If the Skills work the way the announcement suggests, a researcher could ask an agent to run variant predictions and assemble the results without writing the API calls by hand. The caution from our post on AI-designed phage genomes applies here as well: only about 5% of those generated genomes worked, and the lab found out which ones by testing them.

What a business owner can take from this

You will not use AlphaGenome 2, but the design choice behind it carries over. Version 1 answered a general question about a tissue. Version 2 answers a specific question about one cell type, with the lab's own data as context, and its hit rate nearly doubled.

The same thing happens with AI in a small company. A generic chatbot that knows nothing about your price sheet or your customers gives generic answers. Give it your own context and a narrower question, and you get something you can act on.

If you want help finding where that applies in your business, New Face Design offers a free process audit.