NOTE-657: Bhattacharyya et al. 2023, heritable iron memory in E. coli

Contribution

It shows, in single-cell swarm assays, that the heterogeneity in swarming among genetically identical E. coli cells is not noise renewed each generation: a cell’s swarming class is shared by its descendants for about four generations and dissipates by seven. It identifies the cell’s iron status, read through the Fur regulon, as the variable that carries the state, and shows that pushing iron down or up lengthens how long the state is kept. It adds a phenomenological model in which the state is held for a fixed time before switching, rather than switched at a constant rate.

Key insight

A physiological quantity that is diluted, rebuilt and regulated slowly across divisions, here the cell’s iron pool and the Fur-controlled uptake machinery, acts as a short inherited state. It does not need a bistable switch or an epigenetic mark: it is a slow variable, so a lineage keeps the founder’s value for a few divisions and then relaxes back to the population distribution. Pinning the variable (chelator, added iron, a mutant that cannot import) pins the behaviour for longer.

Assumptions

• Strain and conditions. E. coli MG1655; swarm plates of 0.45% Eiken agar with 0.5% glucose, dried 8 h, inoculated with 4 μL and read at 30 h at 30 °C; swarm diameter measured by ruler as the maximum diameter.
• Single cells by dilution to extinction. A suspension targeted at 0.5 cells per 4 μL; Poisson sampling puts a cell on about 40% of plates and more than one on about 5%, which are excluded by counting nucleation centres. At least 120 plates for 50 swarms per assay.
• Generations by timing. G4, G7 and G12 are harvest times in 96-well plates calibrated by CFU counts, not observed divisions; sixteen daughters per mother are expected at G4 and sampled singly at later generations.
• Iron is read, not measured. The biosensor is sfGFP on a plasmid behind the fepA promoter, which Fur represses when iron is high. It reports Fur derepression; plasmid copy-number variation is controlled only by a constitutive Ptrc-sfGFP comparison.
• Model. Equal proliferation rates in all three classes, so fractions can be modelled instead of counts.

Key results

• Fig. 1. Swarm-derived colonies on moist hard agar are less circular and larger than planktonic ones (n = 218, Mann–Whitney P < 0.0001); about 20% of planktonic colonies look swarm-like. Single-cell inocula show significantly higher coefficient of variation in swarm diameter than 100- and 10,000-cell inocula in each of three zone-adjusted comparisons (F-tests), most in the outer zone.
• Fig. 2 (about 1,700 plates). Planktonic G0 mothers vary; the 16 G4 daughters of each mother are homogeneous, and K-means on the column means gives three classes, XS, M and L; by G7 within-sibling variability returns. Swarm-edge mothers are uniform at G0, give one class (L) at G4 with lower noise, and lose it by G7. Mothers from hard agar behave like planktonic ones (Fig. S3B).
• Fig. 3 (about 2,400 plates). No environmental perturbation changes the variability; of the plasmid-expressed genes only fepA and fur do (PCA over mean, SD and noise, PC1 60.06%). DFO shifts the distribution to better swarming, FeCl₃ to worse; neither affects ΔfepA. DFO-treated lineages keep their class to G7 (lost by G12); fepA overexpression keeps the poor-swarming class to G12, the last generation tested.
• Fig. 4. Reporter signal vs swarm diameter in sorted G0 cells: Spearman r = 0.7067, p = 8.66 × 10⁻¹⁶ (WT/PfepA-sfGFP; exponential fit R² = 0.5996); r = 0.1506, p = 0.14 (WT/Ptrc-sfGFP); r = 0.0507, p = 0.5965 (ΔfepA/PfepA-sfGFP). Five sorted mothers per strain followed: in WT/PfepA the G4 daughters follow the mother’s reporter level, and by G7 they do not; ΔfepA, uniformly low in iron, keeps its class to G7.
• SI model. df_XS/dt = −k_{XS→M} f_XS + k_{M→XS} f_M; df_M/dt = k_{XS→M} f_XS − (k_{M→XS} + k_{M→L}) f_M + k_{L→M} f_L; df_L/dt = −k_{L→M} f_L + k_{M→L} f_M, with time in generations. Steady state is fixed at 52% XS, 35% M, 13% L by k_{M→XS} = 1.49 k_{XS→M} and k_{M→L} = 0.37 k_{L→M}. The founding cell’s outgoing rates are zero until τ. Least-squares fit (Excel Solver) to G4 data and the G7 steady state gives τ_XS, τ_M ≈ 3, τ_L ≈ 4 generations, k_{XS→M} ≈ 0.4, k_{L→M} = 1.8 per generation.
• SI Fig. S11. Reporter signal correlates with biofilm (crystal violet; more iron, more biofilm) and with survival after 4 h at about half-MIC kanamycin or chloramphenicol (less iron, better survival), not in the controls. These are G0 measurements only.

Claims

id	claim	strength	support
C1	Genetically identical planktonic E. coli cells differ in swarming potential more than population inocula reveal	strong	Fig. 1, many single-cell assays with F-tests on the coefficient of variation
C2	A cell’s swarming class is shared by its fourth-generation descendants and lost by the seventh, in planktonic and swarm-derived cells alike	moderate	Fig. 2, 25 mothers per condition, three harvest times only
C3	Prior swarming makes the founding cells uniform good swarmers; growth on hard agar does not	moderate	Fig. 2C, E; Fig. S3B
C4	Iron status is the variable that carries the state	moderate	only fepA and fur among ten genes alter variability (Fig. 3A–B); iron manipulation shifts and prolongs it (Fig. 3C–E); reporter correlation (Fig. 4E). Correlational and through a Fur reporter; no direct iron measurement
C5	Daughters’ swarming follows the mother’s iron status at G4	weak to moderate	Fig. 4F–G, five sorted mothers per strain
C6	Constant-rate (memoryless) switching cannot reproduce the lineage data; a fixed holding delay is needed	weak	asserted from the model fit; no fit of the delay-free model is reported, and five parameters are fitted to G4 and G7
C7	Iron status also predicts biofilm formation and antibiotic survival	moderate (G0 correlation)	Fig. S11; inheritance of these was not tested
C8	The memory is a bet-hedging strategy and likely widespread; “iron memory impacts all iron-controlled behaviors”	speculative	Discussion, no test

Method

Dilution to extinction gives single cells, which are either plated directly on swarm agar or grown in 96-well plates to a calibrated generation and their daughters plated singly. Swarm diameter (and, for clustering, total pixel intensity) is the read-out; noise is the coefficient of variation, compared by an F-test. Candidate carriers are screened by environmental and plasmid perturbations and a PCA of mean, SD and noise. Cells are then sorted by the Fur-reporter signal (FACS) into single wells, and each sorted mother’s own swarm, or its G4/G7 daughters’ swarms, are scored against its reporter level. The lineage data are summarised by a three-class linear switching model with class-specific holding delays.

Concepts

• memory: here, any heritable non-genetic carry-over of a cell’s swarming potential across divisions. Used throughout as a functional label, not with a claim about storage and retrieval beyond persistence.
• iron memory: the authors’ name for that carry-over once identified with intracellular iron status as reported by the Fur regulon.
• conditioning: prior swarming, which leaves cells uniformly in the good-swarming class. Borrowed from the vocabulary of associative learning; no stimulus–response association is tested.
• decision-making: the choice of when to start swarming and how well to swarm on a new surface. Not a model of choice among options.
• SCI: single-cell inoculum swarm assay.
• XS, M, L: swarm diameter under 35 mm, 35–65 mm and over 65 mm.

Connections

It builds on the observation, from the Harshey laboratory’s earlier work, that swarm-derived cells start swarming after a shorter lag, and on the known role of iron starvation as a swarming signal and of Fur as the repressor of iron uptake. It places itself among bacterial memory mechanisms that are stimulus-specific (fimbrial phase variation, chemotaxis adaptation, the motile–sessile switch) and claims a different kind, a state reached by several stimuli. It explicitly sets aside bistable and toggle switches and epigenetic inheritance because the memory lasts only a few generations, and invokes stochastic gene-expression noise and dilution of Fe–S proteins at division instead.

Bearing on the record

• New THEORY. The record held no account of non-genetic, multigenerational state in single bacteria from experiment. This reading produces THEORY-179, stating the finding (a slow physiological variable carries a cell’s behavioural state for about four generations), with what the evidence does not reach.
• Against the record’s other bacterial-memory reading. Irreversibility in bacterial regulatory networks (Zhao et al. 2024, filed the same day) argues from Boolean models that a transient perturbation can leave the E. coli regulatory network in a different self-maintaining attractor. This paper is the opposite case, read in cells: a heritable state that is not self-maintaining, decays in about seven generations, and is explicitly not a bistable switch. It is not evidence for that account; it shows that a few generations of inheritance do not by themselves indicate an alternative stable state. My connection, not the paper’s.
• THEORY-136 (a persistent or abrupt change is not evidence of alternative stable states). The same lesson at the cellular scale: the persistence here is explained by a slow variable relaxing, and the authors reject multistability. My connection; it does not change that THEORY.
• The word “memory”. Readings in the record that weigh claims of memory, learning or cognition in organisms without nervous systems should take this paper’s “memory”, “conditioning” and “decision-making” as functional labels for persistence of a physiological state. Nothing in it bears on learning in the sense of the learning-and-conditioning tag.
• No instruction for machine-learning practice; nothing for the anthology.

Limitations

• The iron link is correlational and indirect. No intracellular iron is measured; the reporter reads Fur-repressed transcription from a multicopy plasmid. The strongest evidence (Fig. 4E) is one strain’s correlation over about 96 cells.
• Inheritance rests on few founders and three time points. 25 mothers per condition in Fig. 2, five sorted mothers per strain in Fig. 4, and only G0, G4, G7 (and G12) were sampled, so “lost by the seventh” is “lost at the seventh time point tested”. Generations are inferred from timing, not observed.
• The delay claim is weakly supported. The model has five fitted parameters for data at two time points, with steady-state fractions imposed; the comparison with a delay-free model is asserted, not shown. The authors say three classes may stand in for a continuum.
• Biofilm and antibiotic links are single-generation correlations.
• Some figure-level details (the hard-agar control, the noise per mother, the growth/motility clustering) are in SI figures not examined here.
• The bet-hedging interpretation and the claim that the mechanism is general are not tested.

Open questions

• Does the state follow iron itself? A direct single-cell iron measurement (or a chromosomal reporter of a second Fur target) correlated with swarm outcome would separate iron from Fur-promoter or plasmid effects.
• Is the decay a fixed holding time or a slow continuous relaxation? Lineages sampled at every generation, with the delay-free and delayed models both fitted, would decide it.
• Is swarm-edge “conditioning” more than selection of low-iron cells that happened to reach the edge? Tracking iron status of the same cells before and after swarming would answer it.
• Do biofilm propensity and antibiotic tolerance inherit on the same four-generation timescale?