Google DeepMind Launches AlphaGenome 2 — DNA Predictions Now Down to the Single Cell

A DNA variant doesn’t behave the same way in every cell. It might matter in an immune cell and do nothing at all in a skin cell — which is exactly why averages can mislead. Google DeepMind‘s new AlphaGenome 2, launched on November 11, 2026, is built around that problem: instead of predicting what a DNA sequence does across a broad tissue average, it aims at specific cell types and cell states.
From Tissue Averages to Single Cells
The original AlphaGenome worked at the level of tissues — useful, but blunt. A tissue is a mix of many cell types, and an average across all of them can hide the one cell type where a variant actually does something. AlphaGenome 2, as reported by SXF Brief and The Neuron on November 12, moves the prediction down to specific cell types and cell states. To learn those differences, it was trained on more than 107 million individual cell profiles spanning 68 tissues.
The training data is the other half of the story. One hundred and seven million cell profiles is a lot of single cells, and spreading them across 68 tissues means the model sees the same genome behaving differently in different biological settings. That’s the raw material a cell-level predictor needs: not just what a sequence says, but where it’s being read. Without that spread, “cell-specific” would be a label, not a capability.
Why does that granularity matter? Because disease-linked variants often act in narrow places. A variant tied to an immune condition, for instance, may only change gene activity inside particular immune cells. A tissue-level average could wash that signal out entirely. A cell-level model at least gives researchers a sharper question to test in the lab.
The TenK10K Numbers
DeepMind’s reported benchmark for the launch is the TenK10K immune-cell dataset. On it, AlphaGenome 2 identified 30% of causal expression variants — the variants that actually drive changes in gene expression — against 16% for the original AlphaGenome. Close to double, on the same dataset. That gap is the clearest sign of what the extra resolution buys: put plainly, the model is catching effects the first version missed because they were averaged away.
A quick note on what’s being counted. A causal expression variant is a change in the DNA that actually causes a gene’s activity to rise or fall in a cell — not just one that happens to sit nearby. Telling the causal ones apart from the bystanders is one of the slow, expensive parts of genetics research, and it’s where a better predictor earns its keep: every correctly flagged variant is a lab experiment that starts with a head start.
Worth keeping in proportion, though: 30% also means 70% of causal variants on that dataset weren’t identified. This is a research instrument getting sharper, not a solved problem. Benchmarks on one immune-cell dataset also don’t tell us how the model performs everywhere else — that picture only comes as independent groups put it to work.
Research Only, and DeepMind Is Explicit About It
AlphaGenome 2 is available through the AlphaGenome API, and the terms matter as much as the technology. Access is for non-commercial research only. DeepMind states plainly that the model is not approved for clinical or diagnostic use. So no, this isn’t a tool doctors can use to read a patient’s genome today, and it shouldn’t be reported as one. What it is, for now, is a stronger instrument for the people studying how variants behave — the lab work that has to happen long before anything clinical is even on the table.
A Busy Stretch for Google’s AI Research
AlphaGenome 2 lands in the middle of a crowded run for Google’s AI teams. We recently covered Gemini 4 Carbon topping the MathArena leaderboard and Google open-sourcing XSE, the serving engine behind its search stack. Different fields, same pattern: Google is pushing its models into narrower, harder problems and publishing the numbers as it goes. Whether AlphaGenome 2 holds up under outside scrutiny is the next thing to watch — the API is where that scrutiny will happen.
For researchers, the offer is straightforward: cell-level predictions, a training set of more than 107 million cell profiles behind them, and a benchmark that suggests a real step up from the first model. For everyone else, the honest summary is narrower. A powerful research tool just got sharper — and its maker is being unusually clear about what it isn’t for.
