DeepMind Says AlphaGenome 2 Nearly Doubles Its Variant Hits, but It Stays Out of the Clinic
Google DeepMind says AlphaGenome 2 can predict how DNA changes affect specific human cell types and finds nearly twice as many gene-altering variants as its last model. It is open to non-commercial researchers now, and it is not cleared for clinical or diagnostic use.
Every cell in your body carries the same DNA. A heart muscle cell and an immune cell fighting an infection just read it differently. That is the problem Google DeepMind says its new model, AlphaGenome 2, is built to crack: predicting how a single DNA change plays out in one specific cell type, and barely registers in the one next door.
DeepMind announced the model on November 11 and opened it the same day to non-commercial researchers through the AlphaGenome API, with a private preview for commercial research. Before anyone gets carried away, the company is blunt about the limits. Its announcement post states that AlphaGenome and AlphaGenome 2 are intended for research use only and have not been validated or approved for clinical or diagnostic purposes. This is a lab tool. It does not diagnose anyone.
The original AlphaGenome, released last year, took a million-letter stretch of DNA and predicted thousands of outputs across broad tissues in parallel. DeepMind admits that design spread the model thin and made it hard to separate closely related cell types. Version 2 flips the approach. It is a conditional model that takes one target condition at a time, whether a tissue, a cell type or a cell state, and feeds it in alongside the DNA sequence at every stage of the architecture.
The training pile is big. DeepMind says the model learned from more than 100 million individual cell profiles across 68 tissues, and elsewhere in the same post says it trained on bulk and single-cell data from human and mouse samples, including over 107 million individual cell profiles.
The numbers are DeepMind's own
The headline claim is that AlphaGenome 2 identifies nearly twice as many variants that change gene expression in cells as DeepMind's previous state-of-the-art model. On TenK10K, which DeepMind describes as a large single-cell map of 28 primary immune cell types, the company reports the model caught 30 percent of causal variants versus 16 percent, and 43 percent of directional gene effects versus 23 percent, compared with the original AlphaGenome fine-tuned on single-cell data and other specialized tools. Those are company-reported results, not an independent audit.
DeepMind's showcase example is also worth reading carefully. It says AlphaGenome 2 computationally reproduced a previously published finding that a variant in a distant regulatory switch drives inflammatory bowel disease by overactivating the ETS2 gene in gut macrophages. That is a model matching a known result, not a new discovery.
The more interesting trick may be flexibility. DeepMind says researchers can upload single-cell data from just a few dozen cells and get predictions for cell states the model never saw, without retraining. It can also fill in missing measurements, predicting chromatin accessibility from gene activity, for example, which could help labs decide which experiment to run next. Sarah Teichmann, a co-founder of the Human Cell Atlas consortium, was quoted in DeepMind's own post praising exactly that.
"What makes AlphaGenome 2 exciting is how it adapts directly to new single-cell data without retraining, predicting how DNA variants act in newly profiled cell states from even a small sample of cells." Professor Sarah Teichmann, Human Cell Atlas co-founder
There is a catch DeepMind does not put in its headline. Genomics lead Ziga Avsec told The Verge that AlphaGenome 2 is not replacing its predecessor any time soon, because its more detailed predictions are more expensive to produce. He said that cost would have made it impractical to build the genome-wide Atlas DeepMind released in September, which covers every possible single-letter change in the human genome, with the new model.
Avsec framed the payoff in drug development terms, telling The Verge that drug makers target specific proteins and specific cell types. That is the pitch, and it is a long road from a research API to anything a patient would notice. AI labs keep racing to claim science wins, from Anthropic saying Claude found a never before seen enzyme to DeepMind's growing biology stack. AlphaGenome 2 is a credible entry in that race. Whether its numbers hold up is now a question for the outside researchers who can finally test it.
