Until now, one of the panelists said on Moonshots #288, when a patient showed up with a mutation nobody had seen before, the doctor had to guess whether it mattered.
The arithmetic behind that problem is simple. A human genome is roughly 3.2 billion letters long, in two copies, one from each parent. The letters are A, T, C and G, so at any single position there are three other letters it could have been. Three changes times about three billion positions is more than nine billion possible single-letter variants — and almost all of them have never been seen in a patient, let alone studied.
On 8 September, Google DeepMind published AlphaGenome Atlas, which runs its AlphaGenome model over all of them in advance. The result is about a petabyte of predicted molecular effects: what each substitution is expected to do to the machinery of the cell, in stretches of DNA that code for proteins and in the regulatory DNA that controls when genes switch on. Google pairs the predictions with a single AlphaGenome Variant Impact score, so a researcher can rank candidate variants instead of reading thousands of separate outputs, and puts the whole collection behind a portal that needs no code to query.
"Bulk solving" a field
Alex, one of the panelists, said the release follows a pattern he now expects to repeat. AlphaFold began as a model, then a better model, then AlphaFold 3 and a Nobel Prize; then it became the AlphaFold Protein Structure Database, in which the model was used to precompute the answers for essentially the whole of single-molecule protein structure and give them away. "Bulk solving a field seems to want to become a database of all the pre-computed answers to all the questions that can be asked in that field," he said. Variant effect prediction, in his account, is the same move applied to nine billion possible single-nucleotide changes.
His claim is that this works wherever the problems can be written down as a finite list, even a very long one. He offered mathematics as another case: Paul Erdős left behind a finite set of open problems, and Alex said he might as well call that class "essentially effectively solved at this point" — his own judgment, not a formal result. The host compared the Atlas to Mendeleev's periodic table, which predicted elements before anyone found them, and said DeepMind uses the same analogy.
Another panelist, who said he was one of the authors on the OpenFold paper, made the more practical point about why a precomputed collection differs from a capable model. "Running the models is one thing," he said; "having a completely comprehensive list is another." A laboratory that wants an answer no longer needs the hardware, the pipeline or the expertise to run inference; it needs a query. He expects leaps forward simply from the lowered bar to access.
A namespace for rare conditions
Host Peter Diamandis turned the same fact toward patients and families: if someone in your family has a disease, find other people with the same condition or genetic syndrome, pool capital and fund the work rather than waiting for somebody else.
A panelist extended that into a suggestion about organizing. For the first time, he said, an otherwise idiosyncratic condition may just be "a row in a billion row lookup table." If everyone carrying the same single-nucleotide change can identify it by the same entry in the same public table, they share a name for what they have — "a common namespace, a common lexicon" — and, with it, a natural way to find each other and coordinate, not only to pool money. That is a proposal about what the resource makes possible, not something the Atlas itself provides; Google describes its role as prioritizing variants and interpreting predicted effects for researchers.
"It's not actually a lookup table"
The sharpest qualification in the segment came from a panelist recalling a walk through Cambridge with Noubar Afeyan, the founder of Flagship Pioneering — the venture firm where Moderna was created, and which Afeyan chaired for years. Told that his companion worked on neural nets, Afeyan said that someday there would be very capable ones and that their fundamental use would be mapping genotype to phenotype: going from the genetic sequence to the observable trait or disease. "I was like, wow, that's a very specific sentence," the panelist said. "So here we are. It's that day."
Then he took the framing apart. "We're calling this a lookup table, but it's not actually a lookup table." There is a table, with nine billion starting combinations in it. But the interactions between any two of those genetic switches matter to the outcome, and the table has one row per single change. What the Atlas really is, in his description, is a lookup table that feeds a domain-specific neural network, which predicts the outcome for a particular combination.
That network has to be trained somewhere, on something. It needs local tuning on the relevant dataset and, above all, lots of phenotype or outcome data — records of what actually happened to people or organisms — so that it can interpolate between cases. Which is why he thinks the precomputed table does not end the business of interpreting genomes but multiplies it: every domain becomes a separate opportunity, the outcome data for each one stays scarce, and whoever holds the domain-specific advantage in any single field has, in his view, a sustainable long-term business. He expects that pattern to repeat in thousands of domains.
So the Atlas answers a narrower question than the periodic-table analogy suggests. It predicts what one letter change is expected to do inside a cell, for every letter change, and hands that to anyone with a browser. What a given person's combination of variants will mean for their body is the part still waiting on data nobody has finished collecting.