Google's new DNA model nearly doubles its hit rate on disease-causing mutations. Drug companies can only rent it

Google DeepMind on Wednesday released AlphaGenome 2, an AI model that predicts how small changes in DNA alter the behavior of specific kinds of cells. For drug makers, that question sits at the start of almost every program: which genetic change actually drives a disease, and in which cells.

DeepMind says the new model finds nearly twice as many of those culprit changes as its predecessor on a standard immune-cell test. But companies cannot download it. Commercial users must join a private preview, with no published price, and run every query on Google's servers.

That makes AlphaGenome 2 as much a supplier decision as a science result. Pharmaceutical and biotech buyers get a potentially useful tool. They also take on a vendor whose parent, Alphabet, owns its own drug discovery company, Isomorphic Labs.

What the model does, and what DeepMind claims

Most genetic variants linked to disease sit outside genes, in stretches of DNA that act like dimmer switches. Knowing a switch matters is not enough. Researchers need to know which cell type it affects, such as an immune cell, a gut cell or a neuron, before they can design a drug or an experiment around it.

The first AlphaGenome, released in 2025, made predictions across whole tissues. Version 2 lets a researcher name a specific cell type or state, or upload their own single-cell measurements, and get a prediction for that condition. DeepMind says it can handle new experiments without being retrained.

The company says it trained the system on more than 107 million individual cell profiles from human and mouse samples, covering 68 tissues. It reads a stretch of one million DNA letters at a time.

On TenK10K, a dataset covering 28 types of immune cells, DeepMind reports the model identified 30% of likely disease-causing variants. The original model, adjusted for single-cell data, found 16%. AlphaGenome 2 also correctly predicted whether a variant turns a gene up or down 43% of the time, versus 23%.

DeepMind also showed the model could reproduce, by computer alone, a 2024 laboratory finding. A variant far from a gene called ETS2 raises its activity in gut immune cells, a link to inflammatory bowel disease. In a second example, the model told vaccine-activated immune cells apart from resting ones.

The caveats behind the headline numbers

Every performance figure comes from DeepMind's own paper. No outside group has published a replication yet.

Even taken at face value, the 30% figure means the model still misses roughly seven in ten causal variants on that test, at the strict bar DeepMind set, where nine in ten of its picks must be right. DeepMind's paper itself says the model can underestimate some effects, and that diseased and stimulated cell states are thin in its training data.

The ETS2 result also confirms something scientists already knew. As The Neuron pointed out, the real test is whether the model leads researchers to mechanisms nobody knew about, which then hold up in the lab. That has not been shown yet.

DeepMind says AlphaGenome 2 is for research only and has not been validated or approved for diagnosis or clinical use. The paper also notes it is costlier to run than the original, because each cell condition requires a separate prediction.

Why the access terms matter to buyers

Academic and other non-commercial researchers can use the model free through DeepMind's AlphaGenome API. Commercial researchers must apply to a private preview. DeepMind has not published a price.

The first version's trained model files, known as weights, were released for non-commercial use. Version 2's were not, so it exists only as a hosted service, i-SCOOP and Startup Fortune report.

According to i-SCOOP's reading of DeepMind's FAQ, the model's outputs may not be used to train other AI models. That rules out a common cost-saving move: using a large model's answers to build a cheaper in-house copy.

Researchers who tailor predictions to their own samples must upload those cell measurements to Google. i-SCOOP says the public FAQ does not spell out how that data is stored or used. That is a live question for anyone working with patient-derived material.

Then there is the competitive question. Isomorphic Labs is in the drug discovery business, so some would-be customers are also its competitors. Startup Fortune notes DeepMind has not said how it would handle a customer that competes with its sister company.

Early partners are signing on regardless. Genomics Limited, led by chief executive Peter Donnelly, says it plans to build AlphaGenome 2 into its Mystra platform. Cambridge stem cell scientist Sarah Teichmann, quoted by DeepMind, praised how it adapts to new data "without retraining."

For buyers, the trade-off is now clear. The promise is fewer wasted lab experiments chasing the wrong genetic lead. The price is dependence on a single supplier, with unpublished fees, limits on reusing its answers and drug ambitions of its own.