AlphaGenome 2: What It Means for Your DNA Test

Your genetic test comes back, and one line on the report reads "variant of uncertain significance." Nobody can tell you whether it matters. Your doctor shrugs, the lab says it will update you if anything changes, and you are left holding a letter and a number that means nothing yet.

That gap between reading DNA and understanding it is exactly what AlphaGenome 2, the new model Google DeepMind released on November 11, is built to narrow. It predicts how a single change in your DNA alters gene activity inside one specific type of cell, rather than across a whole blended tissue. That sounds like a small technical shift. It is the difference between knowing a fire alarm went off somewhere in a building and knowing which room it came from.

This article explains what AlphaGenome 2 actually does, what its numbers mean in plain terms, and what it can and cannot change about a DNA test you take today.

Why Most of Your DNA Is Hard to Read

Start with a fact that surprises most people. According to the National Human Genome Research Institute, only about 2% of your genome actually codes for proteins. The other 98% is non-coding DNA.

Non-coding does not mean useless. Much of it works like a panel of dimmer switches, telling genes when to turn on, how strongly, and in which cells. A single letter change in one of those switches can leave a gene's recipe perfectly intact while making the gene run too hot or too cold.

Here is where it gets tricky. Every cell in your body carries the same DNA, but a gut immune cell and a brain cell read it very differently. A switch that matters enormously in one cell type might do nothing in another. So a variant can look harmless in an average of many cells and still cause trouble in the handful of cells where that switch is active.

Clinical labs are much better at judging changes inside protein-coding genes, where the effect on the protein can often be reasoned out. Changes in the regulatory 98% are far harder to call, which is one reason so many results end up in the uncertain pile.

What AlphaGenome 2 Does Differently

The original AlphaGenome predicted many tissues in parallel from a stretch of DNA. AlphaGenome 2 works as what DeepMind calls a conditional model. You hand it a sequence of up to 1 million DNA letters plus one target condition, such as a tissue, a cell type or a cell state, and it predicts what happens in that context.

According to Google DeepMind's announcement, the model was trained on bulk and single-cell data from human and mouse samples, including more than 100 million individual cell profiles across 68 tissues. Three upgrades stand out:

• A cell encoder. Scientists can condition the model on their own single-cell data. DeepMind says a snapshot from just a few dozen cells can be enough, with no retraining.
• Cross-modality imputation. Give it one type of measurement, such as gene activity, and it predicts readouts nobody measured, such as how open the DNA is in that cell.
• DNA methylation. The model now predicts these chemical tags on DNA, which the first version did not.

If you want the underlying idea in one line, it is pattern matching at scale. The model is a form of machine learning that has seen how millions of cells read their DNA, and it uses those patterns to guess how an unseen sequence will be read in a chosen cell.

The Numbers Behind AlphaGenome 2, Translated

DeepMind's headline result comes from TenK10K, a single-cell benchmark covering 28 types of primary immune cells. The task is to find the variants that actually cause a change in gene expression and to say which direction that change goes.

The reported scores:

• Causal variants identified: 30% for AlphaGenome 2 versus 16% for the previous state of the art (the original AlphaGenome fine-tuned on single-cell data, plus other specialized tools).
• Direction of gene effect identified: 43% versus 23%.

A worked example with real numbers

Picture a research team studying an immune disease with 100 true causal variants hiding in its data. The older approach would reliably flag about 16 of them. AlphaGenome 2 would flag about 30.

That is 14 extra leads for the lab to test, an 88% increase in useful candidates from the same data. It also means roughly 70 of the 100 causal variants still go unflagged. Nearly doubling a low score is real progress, but the work of finding what matters is far from finished.

Keep two caveats in mind. These are DeepMind's own benchmark figures, and we could find no independent evaluation of them yet. Its showcase examples were also computational: for inflammatory bowel disease, the model reproduced a published finding that a distant regulatory variant overactivates the ETS2 gene in gut macrophages, and it pinned the effect to those cells. That is a strong sanity check, but it confirms known biology rather than proving a new discovery in the lab.

What It Means for Your DNA Test Results

So what does this change for you right now? Directly, nothing. DeepMind is explicit that AlphaGenome and AlphaGenome 2 are "intended for research use only" and have not been validated or approved for clinical or diagnostic purposes. No lab will rerun your report through it tomorrow.

Indirectly, the stakes are large. A peer-reviewed analysis in Genetics in Medicine looked at 1,512,306 diagnostic tests from 19 North American labs. It found that 32.6% of multi-gene panel results included at least one variant of uncertain significance, compared with 22.5% for exome and genome sequencing.

Panel size drove most of the difference. Small panels of 2 to 10 genes produced an uncertain variant 6.0% of the time. Panels covering more than 200 genes did so 76.2% of the time.

Run the math on your own test

Say 1,000 people take a typical multi-gene panel. Based on that study's rate, about 326 of them get at least one uncertain result. If 1,000 people take the largest panels, about 762 do. That is roughly three out of four people walking away with a question mark, often with real worry attached.

Tools like AlphaGenome 2 matter because they generate evidence about how a variant behaves in the cells relevant to a disease. Over years, that kind of evidence is what lets labs move variants out of the uncertain column, and the study, citing earlier work, notes that many uncertain variants are later reclassified as benign. The model speeds up the research that feeds those reclassifications. It does not hand down verdicts.

Where Researchers and Drug Makers Will Use It First

The near-term users are labs and companies, not patients. AlphaGenome 2 is available now for non-commercial research through the AlphaGenome API, while commercial access is in private preview. DeepMind did not publish pricing.

Drug discovery is the obvious early market. Peter Donnelly, chief executive of Genomics Limited, said in DeepMind's announcement that his company plans to fold AlphaGenome 2 into its Mystra platform for drug discovery. Knowing which cell type a disease variant acts in tells a drug team where a medicine needs to work, which can rule out dead-end targets early. If you have followed what AI drug discovery has actually found so far, this is the same pattern: better maps first, approved medicines much later.

DeepMind also showed the model on vaccine data it had never seen. Given single-cell data from a COVID-19 mRNA vaccine study, it told vaccine-activated T cells apart from resting ones, and it suggested a variant linked to severe COVID-19 disrupts gene activity more strongly in the activated state. That is the cell-state idea in action: the same DNA, read differently depending on what the cell is doing.

There is a deadline for researchers, too. DeepMind's community program takes submissions for featured use cases until November 30, with selected projects eligible for extra quota and credits.

How to Read the Hype Around AlphaGenome 2

New AI science models arrive almost weekly, and the pace of releases makes it hard to tell which ones deserve attention. A few questions help you sort this one, and the next one:

1. Is the benchmark the company's own? Here, yes. Wait for outside groups to test it before treating the gains as settled.
2. Was anything confirmed in a wet lab? The highlighted results so far are computational reproductions and predictions.
3. Who can use it, and for what? Research use only, with commercial access still gated.
4. What is the error rate in plain terms? Finding 30% of causal variants is a big jump and still a minority of the answers.

Run through that list and AlphaGenome 2 lands as a genuine step forward in how science reads the 98% of DNA that does not code for proteins, with the clinical payoff years away.

Frequently Asked Questions

What is AlphaGenome 2?

AlphaGenome 2 is an AI model from Google DeepMind, released on November 11, that predicts how DNA sequences and their variants affect gene activity in a specific tissue, cell type or cell state. It reads up to 1 million DNA letters at a time.

Can AlphaGenome 2 interpret my genetic test results?

No. DeepMind says the model is for research use only and is not validated or approved for clinical or diagnostic use. Talk to your doctor or a genetic counselor about any result on your report.

What does a variant of uncertain significance mean?

It means the lab found a DNA change but does not have enough evidence to call it harmful or harmless. In one large study, 32.6% of multi-gene panel tests included at least one such result, and many are later reclassified as benign.

How is AlphaGenome 2 different from the original AlphaGenome?

The original model predicted many tissues in parallel. AlphaGenome 2 predicts for one chosen cell type or state, can use a researcher's own single-cell data without retraining, and adds DNA methylation predictions.

Is AlphaGenome 2 free to use?

DeepMind has made it available for non-commercial research through the AlphaGenome API but has not published pricing. Commercial access is in private preview.

The Bottom Line

AlphaGenome 2 shifts genome AI from "what does this variant do in a tissue" to "what does it do in this exact cell," and on DeepMind's own benchmark it nearly doubles how many causal variants it can find. That is a meaningful advance for the research that slowly turns uncertain test results into clear answers.

If you are deciding on a genetic test today, the practical lesson comes from the clinical data, not the model: bigger panels produce far more uncertain results. Ask your clinician how many genes the test covers and what the plan is if a question mark comes back. The science reading your DNA is getting sharper, and your best move is to choose a test you will actually be able to act on.