A stimulus that leaves a permanent mark on network state

What the work claims

This is a theory paper: a center-manifold bifurcation analysis plus two levels of simulation, with neurodegenerative disease as the motivating application but no biological experiment. The authors couple a standard susceptible-infected-susceptible (SIS) spreading process, a minimal model in which a "pathology" state hops between connected nodes while nodes recover and become susceptible again, to a continuous neuronal activity variable at each node. The coupling runs both ways: pathology changes the node's excitability, and the node's firing rate changes how strongly it transmits pathology onward.¹

The central claim is that the shape of this feedback, not just its strength, decides the qualitative behavior near the invasion threshold. Nonlinear feedback can produce three regimes that linear-coupled models cannot: forward invasion as in the classical model; finite-amplitude invasion, where the threshold drops away and a sufficiently large seed persists below the nominal epidemic threshold while a small one dies out; and endemic bistability, where two distinct stable prevalence levels coexist at the same transmission rate. Where bistability holds, sweeping the transmission rate up and down traces a hysteresis loop, and the state the network lands in depends on where it has been.¹

How it works

The analytic core is a node-level mean-field reduction of the stochastic SIS process, treated as ordinary differential equations for infection probability and activity, followed by a center-manifold reduction near the epidemic threshold. Three results fall out. First, on general weighted networks, two center-manifold coefficients select between forward invasion, finite-amplitude invasion, and bistability. Second, on regular graphs with homogeneous dynamics the system cannot oscillate, and polynomial coupling functions of degrees m and n give at most mn+1 endemic equilibria; linear couplings in both directions permit finite-amplitude invasion but never endemic bistability. Third, monotone couplings support multistability only when the feedback is reinforcing, meaning both couplings push in the same direction.¹

The authors then check the mathematics against biophysical neuron models at two scales. At population scale they use the exact mean-field reduction of a heterogeneous quadratic integrate-and-fire (QIF) population on a 40-node ring with four connections of weight 1/4 per node, with within-population coupling J = -2, between-population coupling K = 1, mean excitability 0.25, half-width 0.1, and a timescale ratio of 0.05 between neuronal and spreading dynamics. With reinforcing sigmoidal couplings of gains 1.50 and 0.95, the equilibrium branches fold at relative spreading rates of about 0.991 and 1.050, a bistable window straddling the nominal threshold.¹

At single-neuron scale they simulate 1,000 spiking QIF neurons, equivalently theta neurons, scattered on a unit square and connected when closer than 0.155 apart, giving 32,893 undirected edges and mean degree 65.8. Synaptic coupling is 6, the pathology-to-activity gain is 3, the activity-to-transmission gain is 12, and the spreading dynamics run 100 times slower than the neural dynamics so that many spikes occur between infection and recovery events, mirroring the real separation between millisecond spikes and months-long protein spread. At transmission rate 0.90, initial infected fractions of 0.25 and 0.75 settle near mean prevalences of 0.175 and 0.697 respectively: two stable states at identical parameters, differing only in history. A square external input of amplitude 1 applied at every neuron for the interval 500 to 1000 time units pushes the network from the low state to the high one, and a second pulse of the opposite sign from 3000 to 3500 pushes it back. Both states persist after the input is removed.¹

The same pulses move the network across the finite-amplitude invasion boundary at the lower transmission rate of 0.76, where a 50-node seed dies out but a 750-node seed persists. Crucially, when the authors flip the sign of the pathology-to-activity coupling so the feedback is no longer reinforcing, even a five-times stronger input only delays the decay: once the input ends, the infection relaxes to absorption. The persistence is a property of the feedback loop, not of the stimulation itself.¹

Where a skeptic should push

The single most load-bearing assumption is that the SIS process is a defensible cartoon of neurodegeneration. It is not, and the authors know it: real neurons do not recover to a susceptible state after clearing tau, the transmission "network" is plastic and dying rather than fixed, and the pathology-to-activity link in Alzheimer disease runs through specific mechanisms such as activity promoting tau release and amyloid-beta plaques driving nearby neurons hyperactive, not through a generic sigmoid gain. The SIS skeleton is chosen for tractability, which is legitimate mathematics but means every number in the paper is a number about the model, not about tissue. The paper demonstrates internal consistency, not biology.¹

Second, the timescale separation is heroic. The real ratio between a spike and a months-long spreading process is orders of magnitude beyond the factor of 100 the simulation uses, and the paper's "permanent" states are permanent only on the simulated horizon. A skeptical reader should also note that the bistable demonstrations sit in a specific corner of parameter space: gains of 3 and 12 with sigmoid saturation are strong, deliberately reinforcing coupling, and the fold locations at relative rates of 0.991 and 1.050 describe that corner, not a universal margin.¹

Third, the "control" experiment of reversing the coupling sign changes the biological hypothesis, not merely a sign convention. It is the right mathematical control and weaker evidence than a same-sign, weaker-gain comparison would have been. And the spatial network is an arbitrary 2D geometric graph; the authors show mean-field and stochastic levels agree with each other, but both could share the mean-field's blind spots on real connectomes.

The array assay as a bifurcation instrument

Strip the disease framing and the mechanism reads as a statement about any closed perturbation experiment on a living network: if the measured variable feeds back on the thing you are perturbing, the preparation has a state history, and a one-shot perturbation protocol cannot tell you which branch you are on. At fixed transmission rate, identical networks with identical inputs settled at prevalences of 0.175 or 0.697 depending only on initial conditions. That is precisely the situation an MEA disease-model assay creates and then ignores: a culture seeded with pathogenic protein, pharmacologically driven, or chronically stimulated is a network whose activity and pathology co-evolve, and two wells that received the "same" insult on the same day can be sitting in different basins by the time the recording starts.¹

The opportunity is that the microelectrode array is the only instrument class that can run the experiment this paper actually prescribes. Detecting hysteresis requires sweeping a control parameter up and down while continuously recording a state variable with spatial resolution, which is exactly what a closed-loop MEA does: titrate stimulation charge or seeding dose, record the population response on hundreds of channels, and look for a loop in the dose-response curve rather than a point. A hysteresis loop, its width and the location of its folds, is a richer phenotype than any single-dose metric, and it is measurable in days on human-cell cultures in a way the in vivo months-to-years process is not. For drug screening the implication cuts both ways: a compound that appears to block spread may merely be failing to push a culture across its basin boundary, and one that "rescues" prevalence may leave the network stranded on a low branch that a second insult tips over. Both failure modes are invisible to single-endpoint assays and obvious to a ramp protocol.¹

The threat is quieter: stimulation itself becomes a state-writing operation. In the simulation, brief pulses of amplitude 1, applied uniformly for 500 time units, permanently moved the network between endemic regimes, and the moved state outlived the input. Charge-balanced closed-loop stimulation protocols on MEA cultures, the kind used for seizure suppression and excitability tuning, are exactly this kind of uniform transient input. If the culture's activity-pathology feedback is reinforcing, a suppression protocol trained on one branch can shove the culture onto another, and a post-stimulation "washout" period does not certify a return to baseline, because hysteresis means the return path is not the inverse of the forward path. Worse, the sign of the coupling decides whether this happens at all, and the coupling sign is a biological property of the preparation that no amount of stimulation-parameter tuning reveals.¹

For acquisition hardware the requirements follow directly. Ramp protocols demand long-term baseline stability across days, stimulation with logged, charge-accounted pulses precise enough to reconstruct delivered dose, and, most importantly, recording that stays on during and immediately after stimulation, because the state switch happens in the transient. An acquisition chain that blanks its amplifiers during the stimulus pulse and its recharge phase is structurally blind to the moment the network decides which basin to enter. The field has invested heavily in stimulation artefact rejection; this paper is a reason to invest equally in stimulation-window observability, so the assay can watch the fold being crossed rather than infer it afterwards.¹

The bottom line

Established: within the model class analyzed, nonlinear activity feedback generically creates finite-amplitude invasion thresholds, hysteresis, and endemic multistability, and these survive from mean-field analysis down to stochastic spiking simulations of 1,000 QIF neurons, with the persistence of switched states tied to the reinforcing sign of the feedback rather than to the stimulation. Asserted, not demonstrated: that any of this structure exists in living tissue. The clean confirmation would be a history-dependent dose response, swept up and down, in a pathogenically seeded organoid culture on an MEA, with loop width as the readout; the clean break would be monotone, history-independent dose responses. Until one of those is measured, the paper's value to the array community is a design discipline: treat every perturbation assay as a bifurcation experiment, log the delivered dose as carefully as the response, and never let a washout period stand in for a down-sweep.¹