Why AlphaGenome Atlas contains nine billion DNA predictions
Change one letter in a sentence and you might get a typo, a new word, or no useful meaning at all. What happens when you change one letter in DNA?
AlphaGenome Atlas tackles that question at an extraordinary scale. Its headline figure, roughly nine billion possible single-letter changes, comes from surprisingly simple arithmetic: about three billion positions in one human genome, with three alternative letters available at each position.
Google DeepMind announced AlphaGenome Atlas on September 8, 2026. The collection brings together precomputed predictions of molecular effects. Its practical promise is to help researchers decide which changes deserve a closer look.
Why three alternatives make nine billion
DNA uses four letters: A, C, G and T. At a position occupied by A, a single-letter substitution could replace it with C, G or T. Repeat that exercise across roughly three billion positions and you reach roughly nine billion possibilities.
The arithmetic counts one change at a time against a starting sequence. It does not count every combination of changes, or every type of mutation, such as a deletion or a large rearrangement. Nor does it mean nine billion variants have been observed in people or tested in laboratories.
Imagine a huge manuscript and a list of every possible one-letter edit. You would have an enormous set of “what if?” questions. You would still need a way to judge which edits alter the meaning.
Some DNA acts like a volume control
Protein-making instructions occupy only a small part of our DNA. Non-coding DNA includes sequences involved in controlling gene activity: where an instruction is used, when it is used and how strongly.
A useful analogy is a recording studio. Knowing the notes of a song is only part of understanding the performance. Volume controls and timing also matter. Likewise, changing a regulatory sequence can affect a gene without rewriting its protein-making instructions. This analogy does not give every stretch of non-coding DNA a known job; much remains to be understood.
AlphaGenome, the model behind the atlas, predicts molecular readouts such as RNA production and splicing, the cutting and joining of RNA. First announced in June 2025, AlphaGenome’s model paper appeared in Nature on January 28, 2026, months before the Atlas announcement.
What the AI actually compares
Consider an illustrative example. A researcher has an original DNA sequence and another version with one letter changed. AlphaGenome predicts molecular activity for both. Comparing the outputs could suggest that the change reduces RNA production or alters how an RNA message is assembled.
The example is a way to understand the method, not a reported experimental finding. The computer has made a forecast. An experiment can then ask whether cells behave as predicted.
The atlas stores these forecasts in advance. Think of the difference between asking for a single route and opening a map of the surrounding area. Having many results together makes it easier to compare candidates and look for patterns. For a laboratory choosing its next experiment, that could be more useful than another isolated prediction.
What happened when researchers tested a clue
One example involves DNM1. External research collaborators used the atlas’s ranking score to highlight a variant and investigated a predicted change in RNA splicing. The Atlas preprint describes laboratory reporter experiments supporting the splicing effect and identifying nearby variants with related effects.
That is a concrete connection between a computational clue and an experimental result. These collaborators participated in the reported research; their tests do not constitute independent validation of the entire atlas. The interesting achievement is narrower and tangible: a prediction helped point researchers toward a mechanism they could investigate.
A molecular forecast has limits
A predicted molecular effect cannot tell someone their medical future. DeepMind identifies distant regulatory interactions and cell-specific behavior as continuing challenges, and says AlphaGenome has not been validated for personal-genome prediction. Development and environment also matter. Its limitations explanation makes those boundaries clear; AlphaGenome is not validated or approved for clinical use.
The next discovery worth watching is what happens when more researchers follow the map into the laboratory. Which clues hold up? Where do predictions fail? Each answer can improve both our understanding of DNA and the next version of the map. Nine billion predictions make a striking headline. Turning the most informative ones into tested explanations is the work ahead.
Checked October 8, 2026. Update when material model changes or new validation findings change this explanation.