DeepMind taught a model to read DNA one cell at a time, and it still misses most of the story

Good morning,

Take one single letter of DNA. Change it. In one kind of cell, nothing happens. In another kind, a gene switches on or off, and that change can sit at the start of a disease. That is the problem Google DeepMind tried to solve on Wednesday, November 11, with a model called AlphaGenome 2.

The first AlphaGenome read DNA at the level of whole tissues. It gave one answer for a whole patch of the body. That answer was an average, and an average can hide the one cell where the change actually matters. AlphaGenome 2 reads the same DNA one cell type at a time. You give it a stretch of DNA, up to a million letters long, and you tell it which cell you care about. A heart muscle cell. A blood immune cell. A gut cell in a certain state. It predicts what the change does in that cell, and only in that cell.

| TenK10K test, 28 immune cell types | Earlier model | AlphaGenome 2 |

| Causal variants found | 16% | 30% |

| Direction of the gene effect right | 23% | 43% |

DeepMind trained it on more than 107 million individual cell profiles from human and mouse samples, across 68 tissues. It can also take a condition from a scientist's own single-cell experiment, with as few as a few dozen cells, and use it without retraining. The readouts cover gene expression, RNA splicing, chromatin accessibility, histone marks, and a new one, DNA methylation.

The table looks like a big jump, and it is, almost twice as many variants found. But read the numbers the other way too. Even at 30 percent, the model still misses most of the variants that statistics point to as likely causes. DeepMind's own paper says so. A model that finds 30 out of 100 is a better torch, not a finished map.

Access tells the same story. The model is open today for non-commercial research through the AlphaGenome API. Commercial research sits in a private preview. And DeepMind is clear that this is research only. It is not validated or approved for clinical or diagnostic use, so no doctor and no patient should treat a prediction from it as a diagnosis.

It also costs more to run than the first model, so DeepMind is keeping the original in service. There will be no genome-wide atlas from AlphaGenome 2, the way there was for the first one. One careful answer at a time is the price of the finer picture.

That trade feels honest to me. Averages are cheap and they hide things. A model that says "in this cell, yes, and in that cell, no" is slower and harder, and it is the kind of answer biology actually needs.

Thanks for reading, and I will see you tomorrow with the next one.

Cover image: AI-generated