The Grammar of the Gene
It is natural to think of a genetic switch on the model of a light switch: a small thing, attached to a larger thing, that turns the larger thing on or off. This picture is intuitive, it is nearly universal in how the subject gets explained to the public, and it is wrong — or at least so incomplete that it amounts to the same error.
A study published this March in Science, led by Idan Efroni at Hebrew University, Zachary Lippman at Cold Spring Harbor Laboratory, and Madelaine Bartlett at the Sainsbury Laboratory in Cambridge, gives an unusually clean opportunity to say why.
The team compared three hundred fourteen plant genomes across two hundred eighty-four species and found two point three million short, non-coding stretches of DNA — regions that specify no protein of their own, but instead govern when and how a nearby gene gets used. These are called conserved non-coding sequences, or CNSs. Some of them are old enough to predate the split between flowering plants and their non-flowering relatives, which puts the oldest at well over four hundred million years; the broader pattern of endurance the researchers describe, across the whole regulatory system rather than any one sequence, is closer to three hundred million. Either figure is startling. Neither is the interesting part.
The interesting part is what, exactly, has been preserved for all that time. It is not the switch. A switch, held in isolation, does nothing. It is not the gene either, which by itself is only a stretch of code with nowhere to go. What has survived is the combination — the particular fact that this sequence, joined to that gene, is capable of a specific piece of work that neither is capable of alone. This is a small distinction with large consequences, and it happens to be close to a distinction the physicist David Deutsch has spent some years insisting on in an entirely different corner of science, under the name constructor theory: the claim that what a physical arrangement is should be described not by its parts, but by which transformations it makes possible, and which it does not. A constructor, on this view, is whatever enables a task to happen — repeatably, without itself being used up in the process. A CNS paired with its gene is a constructor in exactly this sense. Neither piece is the story. The pairing is.
The study gives three findings, and each one is easier to understand once the switch-and-gene picture has been replaced with the combination picture. First: the physical distance between a CNS and the gene it governs drifts substantially over evolutionary time, while the broader relationship — which neighborhood of the genome belongs near which other neighborhood — holds remarkably still. On the light-switch model this is strange: why would the switch wander down the hall and still work? On the combination model it isn’t strange at all. Distance was never what constituted the arrangement. The relation was, and the relation survived the move.
Second: when a genome is shuffled — and plant genomes are shuffled constantly, by duplication, by rearrangement, by the ordinary violence of deep time — a CNS can end up linked to a different gene than the one it originally served. The tempting description is that the switch got reassigned, kept its old job, moved to a new office. But this description quietly imports the very error under examination. Nothing was reassigned, because there was no independently existing “switch” with a persistent identity apart from what it was paired with. A new combination came into being, and that new combination was, immediately and without further explanation needed, capable of a task the old one was not. The sequence carried over. The capability did not — a fresh one was constructed on the spot, out of an old part meeting a new one.
Third, and this is the finding worth sitting with longest: old CNSs do not simply persist — they get duplicated and modified into new CNSs, so that new regulatory sequences are, more often than not, altered copies of ancient ones rather than sequences invented from nothing. Lippman’s own description of this was direct: new regulatory sequences typically come from old ones, modified after the gene they sit beside has been duplicated. In Deutsch’s vocabulary, this is a constructor building further constructors — combinations giving rise to new combinations, each capable of its own distinct task, none of them explicable by an inventory of atoms. What is being copied, across four hundred million years, is not a molecule. It is a piece of working arithmetic: the fact that a sequence of this shape, set beside a gene of that kind, yields a particular outcome and not some other one. The arithmetic gets copied and lightly rewritten. The molecule is incidental to it.
None of this required anything resembling intention on the plant’s part, and it is worth being exact about that, since the vocabulary of “switches doing jobs” tends to smuggle purpose in through the back door. Nothing here wanted anything. A combination either enables a transformation or it does not; when genomes are reshuffled at random and a new combination happens to enable something useful, it persists, and when it doesn’t, it is simply one more discarded arrangement among the great majority that lead nowhere. Anat Hendelman, a co-first author on the study, put the empirical version of this plainly: picking the sequences apart and editing them directly confirmed they are load-bearing for how a plant actually develops, not decorative. That is a claim about mechanism, checked by removing pieces and watching what breaks. It is not a claim about anyone’s plan.
A gene is a noun. A switch is a noun. Four hundred million years did not preserve either noun as such — genes get duplicated, moved, deleted; the switches attached to them drift and rewire constantly, as the study itself shows in detail. What has lasted is the grammar connecting them: the relation that turns two otherwise inert nouns into a sentence capable of doing something. Species have come and gone in that time. Whole floras have been replaced. The grammar, evidently, is harder to kill than anything written in it.
Sources: ScienceDaily, March 2026 · Cold Spring Harbor Laboratory, "Secrets of evolution found in ancient plant DNA" · phys.org, March 2026 · Efroni, Lippman, Bartlett, et al., "A deep-time landscape of plant cis-regulatory sequence evolution," Science, March 12, 2026