Robots Are Better Than People home
Subscribe
Genomics researcher studying single-cell cluster plots and DNA tracks on a lab monitor

AlphaGenome 2 Solved in 30 Minutes What Took a Lab 18 Months

A gene variant behind inflammatory bowel disease took one research team a year and a half of experiments to pin down. Google DeepMind says its new model found the same answer from sequence data alone. Here is what AlphaGenome 2 does and where caution still applies.

In 2024, researchers at the Francis Crick Institute in London identified a new driver of inflammatory bowel disease (IBD): a DNA variant in a "gene desert," far from any protein-coding gene, that switches on a gene called ETS2 in immune cells. It took gene editing and a long series of lab experiments.

On November 11, 2026, Google DeepMind announced AlphaGenome 2 and said the model reproduced that finding computationally. James Lee, the Crick immunologist and geneticist who led the 2024 study, told Nature: "They've managed to figure out the answer in 30 minutes, and it took us a year and a half of experiments." He added: "It's reassuring from our perspective."

The Problem It Targets: The 98% Nobody Can Read

Nature notes that non-coding DNA sequences, the parts that do not spell out proteins but control when genes switch on and off, make up 98% of the human genome and "are among the hardest to decipher." Many disease-linked variants sit there, often in distant regulatory switches called enhancers.

The first AlphaGenome, released last year, predicted variant effects across broad tissues. DeepMind later released AlphaGenome Atlas, predictions for all nine billion possible single-letter changes in the human genome. The catch, as Lee told Nature, was cell type: "Having something that cannot predict cell-type specificity is a bit of a problem, which is why it hasn't been hugely useful for us."

What Changed in Version 2

DeepMind says it redesigned the model to predict for one condition at a time (a tissue, cell type or state) instead of everything in parallel. According to the company's announcement, it was trained on bulk and single-cell data from human and mouse samples, "including over 107 million individual cell profiles," across 68 tissues. Nature reports it covers 637 human and 311 mouse cell types.

AlphaGenome (2025)AlphaGenome 2 (2026)
Prediction unitBroad tissues and cell lines, in parallelOne specific cell type or state at a time
Training dataBulk samplesBulk plus 107 million+ single-cell profiles
New cell statesNot supportedConditions on new single-cell data without retraining
ReadoutsExpression, chromatin, histone marksAdds DNA methylation

On the TenK10K dataset, which maps 28 immune cell types, DeepMind reports AlphaGenome 2 identified "nearly twice as many causal variants (30% vs. 16%) and directional gene effects (43% vs. 23%)" as the original model fine-tuned on single-cell data and other specialized tools.

Field note: In the IBD test, DeepMind says the model "singled out the causal variant and showed that it selectively boosts the ETS2 gene in gut macrophages, leaving other cell types and genes largely unaffected." That cell-specific answer is exactly what version 1 could not give.

New Cells, No Retraining

The feature researchers seem most excited about is adaptation. DeepMind says scientists can upload single-cell data "from just a few dozen cells," and the model can then make predictions for that new state. In one example, using data from a COVID-19 mRNA vaccine study, the company says the model distinguished vaccine-activated T cells from resting ones and showed that a variant linked to severe COVID-19 disrupts gene activity more strongly in the activated state.

Sarah Teichmann, co-founder of the Human Cell Atlas consortium, is quoted by DeepMind: "What makes AlphaGenome 2 exciting is how it adapts directly to new single-cell data without retraining."

The Messy Middle: A Prediction Is Not a Proof

This is where some caution belongs. Reproducing a known finding is a strong check, but it is a check on a result the field already had. The real test is whether predictions on unknown variants hold up in the lab. DeepMind's own language is measured: it talks about helping researchers "prioritize which experiments to do next." Nature's coverage notes the model description is in a preprint, not yet a peer-reviewed paper.

Access is also staged. DeepMind says AlphaGenome 2 is available now "for non-commercial research via the AlphaGenome API, and in private preview for commercial research via Google Cloud."

Decision Matrix: Who Should Pay Attention

  • If you research a disease with known GWAS hits: the precomputed variant scores could narrow your candidate list fast. Validate before you publish.
  • If you work in drug discovery: Genomics Limited CEO Peter Donnelly told DeepMind his company plans to incorporate it. Expect competitors to follow.
  • If you are a patient: this does not change your care today. It may change how quickly researchers find targets over the next few years.

The Inflection Point

Lab biology has always been limited by how many experiments people can run. AlphaGenome 2 does not remove the bench, but it changes the order of work: compute first, then test the few predictions most likely to matter. Turning eighteen months into thirty minutes, even on a known answer, is the kind of ratio that changes how research is funded and planned. The responsibility now is the same as it was for every powerful tool before it: verify, disclose limits, and keep people's genetic data under their control.

The bottom line: AlphaGenome 2 predicts DNA variant effects in specific cell types, found about twice the causal variants of its predecessor on one benchmark, and reproduced an 18-month IBD discovery in 30 minutes. It is a research tool, with predictions that still need lab confirmation.

Sources

We checked every fact, figure and quote in this story against the sources below. Quotes are reproduced as published. Images are AI-generated illustrations, not photos of the events described.

  1. Google DeepMind: AlphaGenome 2, deciphering how cells read DNA to better understand disease (Nov 11, 2026)
  2. Nature News: DeepMind's ultra-powerful AI for genome mapping wows scientists (Nov 11, 2026)
  3. Stankey et al., Nature 630, 447 to 456 (2024): the original ETS2 IBD study
  4. The Verge: Google's AlphaGenome 2 zooms in on the cellular roots of disease

RABTP Newsroom

Every story is researched against primary sources, then reviewed by publisher Sean Hakes before it goes live. We cover threats from the defender's side and never publish exploit details. Read our standards or send a correction.

Keep reading

The White-Hat Brief

Weekly: what the machines learned, and how to stay ahead

One email a week. Verified stories, plain-language defenses, no hype. Unsubscribe anytime.