Two of AlphaGenome 2's authors say the model is almost always right when it says a DNA variant does something, and misses most of the others; what decides a prediction is whether that stretch of DNA is open rather than how far the variant sits from the gene; and the virtual cell is far enough off that they expect AGI first.
A GWAS leaves three questions, and AlphaGenome 2 was built for the third
Avsec says reading a genome-wide association study means answering three questions: which of the dozen or so variants inherited together is the cause, which gene it acts on, and in which cell type it acts. Gayoso says models of this kind used to predict at the bulk level, grinding up a piece of tissue and speaking about the heart or the lung; AlphaGenome 2 is meant to work cell type by cell type, down to a state such as an immune cell that has turned pro-inflammatory. That rests, he says, on ten years of single-cell data, with most of the credit to the Human Cell Atlas.
Distance from the gene is not the problem; DNA wound up tight is
Avsec says people thought these models were particularly bad when a variant sits far from the gene, and the AlphaGenome 2 paper showed that is not quite right: a variant in an accessible region is predicted well even from far away, and one in a region that is not accessible is predicted badly. He compares DNA to a bowl of tangled spaghetti, and says why the tightly packed parts are harder is still a question for their molecular-biology colleagues. One case in the paper: in gut macrophages the model matched a variant to a gene 290,000 bases away, a link a group at the Crick Institute had already validated.
What it flags is almost always real, it misses a lot, and so experiments matter more than ever
Avsec calls these high precision, low recall tools: when the model predicts a strong effect there is almost certainly something there, but there are many false negatives and it misses a lot. Gayoso says a prediction is upstream of the whole investigation a biologist then has to run. Avsec says the models do not really simulate what happens inside a cell and will miss many edge cases, so performing experiments is more important than ever; what the models give is scale and a fast iteration cycle. A lab can plug its own data into AlphaGenome 2 and get variant-effect predictions against that data with no fine-tuning, Gayoso says; the models are free through a Python API and the Atlas is a no-code website.
The virtual cell is far off, and they expect AGI first
Avsec says the bigger goal is the virtual cell Demis Hassabis likes to talk about: give a system a cell state and its genome, apply a perturbation, genetic or a drug, and have it predict how the state changes. He calls it compelling, very hard and still very far off, with the genome a good place to start because only about 10% of drugs pass clinical trials, mostly because we do not understand what causes disease. Gayoso says he is quietly confident the virtual cell can be reached, but the data infrastructure is still being built, and declines to give a number of years. One of the two then says AGI will arrive before a virtual cell; the other agrees, saying intelligence is less hard than biology.
The figures are all in the paper, and the show is the company's own channel
This is Google DeepMind's own podcast, hosted by Hannah Fry, so it is the company talking about its own model, and no figure in it is absent from the AlphaGenome 2 paper, which sits on DeepMind's own site without peer review. The rare-disease case described, a baby diagnosed with epilepsy at six weeks whose variant went unexplained for years until the Atlas score ranked a splicing variant first and a minigene test in cell lines confirmed it, is the DNM1 case DeepMind published with the Atlas in September. The Crick validation was retrospective. Avsec says himself that the Atlas score is a noisy signal and that these tools are not for clinical use.