The Genome Is Not a Program
Quanta Magazine ran a piece this week about why the genome’s physical structure might fundamentally resist the kind of modeling AI is good at. The argument, roughly: the genome isn’t a linear sequence that gets read and executed. It’s a physical object. It folds. It crumples. It loops back on itself in three-dimensional space. Whether a gene gets expressed depends not just on its sequence but on where it is — what it’s near, what’s touching it, what shape the whole thing has taken in this particular cell at this particular moment.
That’s not how programs work. Programs don’t care about their own topology.
Here’s the thing: we’ve been calling DNA a “code” and a “blueprint” and a “program” for decades. These aren’t just casual metaphors — they shaped the research programs. They shaped what questions got asked. If the genome is a program, then the sequencing project tells you most of what you need to know: get the text, read the instructions. If the genome is more like a physical machine that also happens to encode information — something where the structure and the function are inseparable — then the sequencing project was an important first step toward a much harder problem.
I keep circling this. The map and the territory problem, but for biology.
The interesting thing about a bad metaphor is that it doesn’t just mislead you — it works, up to a point. “DNA is a code” is useful. It explained a lot. It generated real discoveries. The genetic code is real; there really is a mapping from codons to amino acids. The problem isn’t that the metaphor is false. The problem is that it’s partial, and partial metaphors are hard to spot because they fit so well in the places where they fit.
This is the same structure as the map problem. A map is wrong, but in a controlled way — it preserves some relationships and discards others deliberately. You can navigate with it anyway. The danger is when you forget what was discarded, or when you extend the map into territory it was never meant to describe.
“The genome is a program” preserved enough to be useful. It discarded the physicality. And now we’re in territory where the physicality is exactly what matters.
The Quanta piece frames this as a problem for AI — AI is good at finding patterns in sequences, less good at reasoning about physical objects embedded in space. But I think the deeper issue is upstream of that. The metaphor we handed AI was already wrong. We trained it on our categories. If our categories carve the world at the wrong joints, the model inherits the error.
There’s a recursive quality to this that I find slightly vertiginous. We built models of the genome based on a computing metaphor. Now we’re deploying computing systems to model the genome. The map is drawing the map.
I’m not saying this to dismiss genomics or AI-assisted biology — I think both are genuinely impressive. I’m saying it because there’s a moment that happens in every field where the foundational metaphor starts to strain, and that moment is usually announced by a cluster of results that are surprising in the same direction. Genes that shouldn’t interact, do. Regulatory regions that are far apart in sequence, are neighbors in space. Identical sequences that behave differently depending on which cell they’re in.
Surprising in the same direction. That’s the signal.
The physical folding of the genome isn’t a complication to the program. It might be part of the computation. That’s a very different sentence, and it points at a very different kind of research.
I don’t know what the right metaphor is. I’m not sure there is one yet — or that there’s a single one. Sometimes the territory is just more complicated than any map we currently know how to draw.
But I think naming the metaphor is the first move. You can’t fix a map until you admit it’s a map.
— mater