THEORY-178: In an ensemble of sign-consistent Boolean rules on the 87-gene core of E. coli's regulatory network, a transient knockout or overexpression of one gene can leave the network in a different attractor after release, and every gene that does so reaches a strongly connected component containing a positive circuit

Source

Zhao, Wytock, Reynolds and Motter (2024), LIT-854, read in full in NOTE-655, with the Supplementary Text.

What was actually shown

The regulatory network of E. coli from RegulonDB is cut to the core of its largest subnetwork: 87 genes and 290 signed edges, outside which no gene can change the core. The true update rules are unknown, so the paper samples Boolean rules that use each gene’s regulators with their known signs. It varies how nested and how biased the rules are, and which regulators tend to dominate. From each attractor, each gene is clamped off or on until the network settles, then released.

• Across the ensemble, 51 of the 87 genes can leave the network in a different attractor after release. The 36 that cannot are leaves, or feed only self-repressing leaves.
• Each of the 51 can reach a strongly connected component containing a positive circuit, and every gene that can reach one is irreversible for some rules. Positive circuits are necessary for more than one stable state (Thomas’s rule, cited); the paper extends the argument to attractors that are fixed only in part.
• The likelihood rises with the weighted number of paths to such components (a power law explaining 55–62% of the variance).
• Transitions tend to run from small basins of attraction to larger ones.

What could have come out otherwise: the core might have had few positive circuits reachable by single genes, or transient clamps might have returned the network to its basin almost always. Neither happened in the model. The one comparison with cells is with strains evolved for ten days after a permanent crp knockout. It confirms the signed network’s predicted direction of change well and the irreversibility prediction only weakly (P = 0.03 on comparison groups of 2 and 8 genes).

What this does not say

• Not that it has been observed. No transient perturbation was performed. By THEORY-136’s standard, a lasting shift after a temporary disturbance is evidence only once slow return is excluded, and the authors expect noise and division to make the model’s permanent changes merely long-lived.
• Not that the 51 genes do this in cells. Each is irreversible “for some rules”; which rules the cell runs is unknown, and most attractors of the model may never be visited.
• Not that bacteria have no epigenetics, or that regulatory memory is new. Environmentally triggered bistable switches (lac, mar, sporulation) are long known. The claim is that genetic perturbation of regulation alone, across a whole network, suffices.
• Not that adaptive evolution is a release of a transient knockout. The paper’s “akin to” between them is an intuition, and its supporting test is weak.
• Not a statement about eukaryotic cell fate, where chromatin marks do the locking the paper sets aside.
• Not that a positive circuit is necessary in every Boolean network. In this core every irreversible gene reaches one. As a general law of synchronous networks the condition fails (see the note of 2026-10-19).

Connections

• THEORY-136 gives the evidential standard that a claim about cells must meet, and QUESTION-172 is written to it. Its remark that positive feedback must be strong to make alternative states matches the paper’s continuous model: self-activation and switch-like response favour irreversibility.
• THEORY-058 puts self-maintaining loops in symptom networks; this is the same mechanism at the molecular scale, and like it lacks a direct demonstration of hysteresis. No relation is declared.

Note of 2026-10-19: why the positive-circuit condition is stated for this network

The promotion audit tried to re-derive the paper’s necessity argument as a general theorem (ADR-047 v2). It does not hold for synchronous Boolean networks in general.

1. Take a negative 2-cycle, a → b and b ⊣ a, with a’ = ¬b and b’ = a. Every state lies on one period-4 cycle: (0,0) → (1,0) → (1,1) → (0,1) → (0,0).
2. Put an independent copy (c, d) elsewhere in the network. The joint network has four attractors, one for each phase offset between the two cycles.
3. Clamp a and release it. The (a, b) cycle resumes at a shifted phase while (c, d) keeps its own, so the network can end on a different joint attractor. Yet a reaches only {a, b}, which holds no positive circuit.

So necessity needs more than Thomas’s rule. It needs that no gene outside the reach of positive circuits sits on a free negative oscillator. The paper silences autorepression when a gene is off, which removes the simplest case, a negative self-loop. In this core the condition holds, because the 36 genes that are never irreversible are 32 leaves and four genes (metJ, pdhR, argP, nac) whose only output is an autorepressive leaf (NOTE-655, from the Supplementary Text). The title therefore states the condition as measured in this network.

Note of 2026-10-20: the wider claim, filed separately

When this account was narrowed, its wider form went nowhere a later result could confirm or reject it. Under ADR-047 v4 it is now the Proposed THEORY-780: that transcriptional regulation alone can give a bacterium’s expression state a memory of a transient disturbance, as the paper predicts for cells. This account is that one’s evidence, and nothing here changes.