AlphaGenome Aims to Read the 97% of DNA That Isn't Genes. The Question Stays Open.

The operating instructions of a human cell sit outside the protein-coding portion of the genome, in a stretch of DNA biology has had the hardest time reading. That stretch holds regulatory switches, structural anchors, splicing signals, the caps that protect chromosome ends, and the sequences that package DNA inside the cell. A single base swap somewhere in that vast region can rewrite a cell's behavior, and the field has been trying to read function from sequence for decades.

The protein-coding portion is "less than 3 percent," as Ars Technica reports. The rest is the part of the genome where biology has had the hardest time reading function, the regulatory layer that runs the cell but rarely shows up in textbooks.

Google wants AlphaGenome Atlas framed as a tool that can now evaluate every possible one-base change across that space, scoring the consequences of 9 billion variants. The honest read is narrower. Ars Technica's piece carries the source's own skepticism directly: it will not be clear what AlphaGenome offers beyond its training data until heavy biological use. That is the pattern under the launch announcement. A model trained to predict function from sequence will pattern-match its labels until its predictions diverge from what the same data already implied.

The mechanism here is the gap between "the model scores it" and "we know what it does." Every "AI decodes the genome" claim runs into the same wall: the tool can only see the correlations its training set contained. Selection will look like causation until the predictions land in a wet lab and survive. If they do, the 97 percent becomes legible and drug targets hiding in regulatory DNA come into view. If they do not, the largest unlit room in biology stays unlit, and the next model inherits the same ceiling.