In Google DeepMind’s rare-disease example, a change in the gene DNM1 stood out for a specific reason: the model predicted that it would tell a cell to cut and rejoin a genetic message in the wrong place. Researchers prioritized that variant, and experimental work supported the predicted mechanism.
That is the practical promise of AlphaGenome Atlas, announced on 8 September 2026. DeepMind has precomputed predicted molecular effects for approximately nine billion possible single-letter changes in human DNA. A researcher investigating an unfamiliar variant can start with an existing prediction and a possible explanation, rather than having to run the model from scratch.
On Moonshots with Peter Diamandis, the panel saw a familiar strategy: use an AI model to build a shared database, then let other researchers build on it. The discussion turned on what becomes easier when predictions are available by lookup—and what remains between that lookup and helping a patient.
The arithmetic behind nine billion
DNA uses an alphabet of four letters: A, T, C and G. At each position, the existing letter can be replaced by one of the other three. Across roughly three billion positions in a human reference genome, that gives approximately nine billion possible single-letter substitutions. The count concerns changes at individual positions, not all the combinations a person’s genome might contain.
Atlas stores about a petabyte of predicted molecular effects, with access through a visual portal and an API, an interface that software can use to retrieve results. Its AlphaGenome Variant Impact score combines information from AlphaGenome with AlphaMissense, DeepMind’s model for assessing substitutions that change a protein’s building blocks. Researchers can use the combined score to prioritize variants across both protein-coding DNA and regulatory DNA, which helps control when and how strongly genes are used.
The resource also offers indications of what drives a prediction: for example, an expected change in gene activity or RNA splicing. Splicing is the process of cutting and joining a gene’s RNA message before it is used to make a protein.
In the DNM1 example, the predicted creation of an incorrect splice site gave researchers a mechanism to investigate. The experimental support concerned that mechanism in that case; it was not a demonstration that every Atlas prediction is correct or that the resource supplies a treatment. DeepMind presents Atlas as a baseline that future models can improve and researchers can use to target experiments.
Pay the computing cost once
The panelist addressed as Alex called the broader approach “bulk solving”: enumerate a field’s questions, compute predictions in advance and turn the results into a database. He compared Atlas with the AlphaFold Protein Structure Database, which made roughly 200 million predicted protein structures available for others to use.
The useful parallel is the distribution strategy. Instead of every laboratory repeating the same large computation, a shared resource makes the results reusable. Neither a predicted protein structure nor a predicted variant effect removes the need to investigate how biology behaves in practice.
Emad Mostaque put the access argument plainly: “running the models is one thing; having a completely comprehensive list is another.” Precomputation, he argued, lowers the barrier to using the work. A smaller laboratory need not run a model across the whole genome just to retrieve predictions for the variants it wants to study.
From genetic letters to biological outcomes
Dave Blundin complicated the lookup-table metaphor with a memory of walking through Cambridge with Noubar Afeyan, founder of Flagship Pioneering, where Moderna originated. Afeyan had told him that a fundamental use of powerful neural networks would be “genotype to phenotype mapping.”
Genotype means genetic makeup; phenotype means observable traits or biological outcomes. Predicting the connection requires more than listing individual DNA changes. Variants can interact, so an inventory of single-letter substitutions is not an exhaustive map of their combined effects.
Blundin described an opportunity for specialized neural networks that use such tables alongside substantial amounts of phenotype or outcome data. His argument was about what entrepreneurs could build on top of shared predictions, not a description of an established Atlas system that already predicts every combination. He sees a durable business opportunity in domain-specific models and the data needed to train them.
That leaves a role for proprietary outcome datasets even when the underlying predictions are widely accessible. Atlas does not itself supply all the observations needed to connect a variant with what happened to a person, establish a diagnosis or determine which intervention would help.
A rallying point for patient research
Peter Diamandis urged families facing a disease to find others with the same condition, pool capital and fund research. Another panelist suggested that a shared variant entry could provide “a common namespace, a common lexicon”—a way for scattered patients to recognize a possible connection and organize around it.
That is a coordination proposal, not a patient-matching or treatment service announced with Atlas. Sharing a variant does not by itself establish that it caused everyone’s condition, and the table is not a catalog of diseases with one answer per row.
For a patient group working with researchers, however, an entry could give a funded investigation a more specific starting question. The DNM1 example shows what that can look like: not simply asking whether a DNA letter matters, but testing whether it creates a faulty splice signal. The lookup supplies a candidate explanation; the next work happens in the experiment.