betaFeedback

WorksAgricultural and Biological Sciences

Hysteresis and multistability in network spreading with neuronal activity feedback

Abstract

Spreading processes on networks often interact with other dynamics on the same nodes. Neurodegenerative disease provides one example: pathological proteins spread through anatomical connections, while neuronal activity influences and is altered by their spread, forming a spreading-activity feedback loop. However, models coupling pathological protein spreading and neuronal activity have largely focused on linear feedback between the two processes. Here we show that nonlinear feedback can fundamentally change the invasion dynamics in a susceptible–infected–susceptible spreading process coupled to a co-evolving activity process. On general weighted networks, the strength and shape of the feedback may generate finite-amplitude invasion thresholds, hysteresis, and endemic multistability. On regular graphs with homogeneous dynamics, we rule out periodic solutions and show that the degrees of polynomial coupling functions bound the number of endemic states, while monotone couplings require reinforcing feedback for multistability. We test these predictions in simulations of a stochastic spiking neuronal network described by quadratic integrate-and-fire dynamics, where we recover both finite-amplitude invasion thresholds and endemic bistability. These results show that feedback from activity processes can create hysteresis and multistability in network spreading dynamics. In neuroscientific applications, our work suggests that neuronal dynamics may act as a control point in neurodegenerative disease, with even transient changes in activity capable of tipping the brain between health and disease.

Introduction

Spreading processes on networks describe a wide range of phenomena including infectious disease, social contagion, and even the progression of neurodegenerative disease. In the simplest description, spreading occurs over a fixed network independently of any other process. In many systems, however, spreading is coupled to a second process that evolves on the same network. During an epidemic, prevalence may change contact behavior, while those behavioral changes affect future transmission [1, 2, 3, 4, 5]. Disease burden may deplete resources for prevention or recovery, while resource availability determines how effectively further spread is controlled [6]. On online platforms, information diffusion can create new social connections, while those connections shape subsequent diffusion [7]. Each of these examples forms a feedback loop in which spreading changes a second dynamical process, which in turn changes subsequent spreading. Neurodegenerative disease provides another key example, with feedback between pathological protein spreading and neuronal activity.

Many neurodegenerative diseases, such as Alzheimer’s disease, involve the accumulation of pathological proteins, which disrupt neuronal function and contribute to disease progression. These proteins can spread through anatomical connections between brain regions, so that disease progression can be viewed as a spreading process on the brain’s anatomical network [8, 9]. Moreover, protein spreading influences—and is influenced by—neuronal activity, forming a feedback loop. In Alzheimer’s disease, for example, neuronal activity can promote the release and spread of the pathological protein tau [10, 11], and amyloid-β\beta plaques can make nearby neurons hyperactive [12]. Although several network models describe pathological protein spreading on brain networks [13, 14], few have considered its coevolution with neuronal dynamics [15, 16, 17, 18]. Furthermore, models describing the coevolution of pathology spreading and neuronal dynamics have largely used linear coupling between the two processes, leaving the consequences of nonlinear transmission feedback comparatively unexplored [16, 18].

The susceptible–infected–susceptible (SIS) process is a minimal model of network spreading and provides a natural starting point for studying such feedback. In the SIS model, infection spreads between connected nodes, while infected nodes recover and become susceptible again. Under common mean-field approximations, an epidemic state and a disease-free state switch stability through a transcritical bifurcation when the transmission-to-clearance ratio crosses a threshold determined by the spectral radius of the network [19, 20, 21]. The threshold determines the onset of invasion, beyond which we observe epidemic spreading throughout the network. However, extensions beyond the canonical SIS model such as nonlinear contact or treatment, higher-order interactions, and adaptive feedback can produce a wider range of dynamical phenomena including backward bifurcations, hysteresis, discontinuous transitions, additional endemic states, and sustained oscillations [22, 1, 23, 24, 25, 26, 27, 2, 3, 4, 5]. Therefore, we may expect similar dynamical phenomena to arise from feedback loops between spreading and neuronal-dynamical processes.

Motivated by neurodegenerative disease, we couple SIS spreading to general activity dynamics and investigate how feedback between the two processes shapes the bifurcation structure near the invasion threshold. Although our main application is neuronal, we initially keep the activity subsystem general, assuming only that it has an attracting fixed point. The coupling acts locally at each node: the spreading state influences the node’s activity, while its activity modulates its outward transmission strength. In contrast to earlier work, we consider general feedback interactions, including nonlinear ones, and determine how they alter the onset and nature of invasion.

We first study invasion near the epidemic threshold on a general weighted graph. To make the analysis tractable, we use a node-level mean-field approximation of the stochastic SIS process to formulate the coupled network dynamics as a system of ordinary differential equations. Center manifold reduction shows that feedback between spreading and activity produces three qualitatively different regimes: forward invasion, finite-amplitude invasion below threshold, and bistability between low- and high-endemic states above threshold. The bistable regimes produce hysteresis, as increasing and decreasing infection transmission rates can lead the system along different stable branches. To gain intuition in a simpler setting, we consider homogeneous dynamics on regular graphs with a minimal activity model, where the full nonlinear system reduces to two variables. The homogeneous system cannot sustain periodic oscillations, and polynomial feedback couplings of degrees mm and nn give at most mn+1mn + 1 endemic equilibria, even when the activity subsystem has a unique stable state. More generally, monotone couplings can support multistability only when the feedback between spreading and activity is reinforcing, and linear couplings may only support finite-amplitude invasion and no endemic bistability. Finally, we test whether finite-amplitude invasion and endemic bistability are present in numerical simulations with biophysical quadratic integrate-and-fire (QIF) neuronal models. We study both population-level dynamics and networks of individual spiking neurons coupled to a stochastic SIS process. Numerical simulations at both levels reproduce the same multistable dynamics predicted by the center-manifold calculation and the regular-graph reduction.

This article is organized as follows. We first study the transcritical bifurcation on a general graph with the coupled SIS-activity system and identify the center-manifold coefficients that control invasion. We then study a minimal coupled system on regular graphs to investigate the global nullcline geometry and the number of homogeneous endemic states. We subsequently validate these regimes in biophysical neuronal models.

Invasion near the epidemic threshold on general networks

We begin by asking how feedback between spreading and activity changes invasion near the epidemic threshold. To address this question, we study an SIS spreading process on a weighted network that coevolves with activity dynamics at the nodes. In the underlying stochastic formulation, the SIS states change through discrete infection and clearance events, whereas the activity variables evolve continuously according to ordinary differential equations. We then use a node-level microscopic Markov-chain approach (MMCA) for the SIS process [28, 29, 30] and replace the stochastic input to activity by its mean-field value. This gives a coupled system of ordinary differential equations for the node infection probabilities and activity, on which we perform the bifurcation analysis. Although motivated by pathological protein spreading and neuronal activity, no particular activity model is assumed. We compute the epidemic threshold and perform center manifold reduction to investigate the impact of the spreading-activity feedback on the bifurcation structure near the threshold. We find forward invasion, finite-amplitude invasion below threshold, and bistability between endemic states above threshold.

Continuous-time infection and activity dynamics

Let Xi(t)∈{0,1}X_i(t) \in\{0,1\}, i=1,…,Ni = 1,\ldots,N, denote the stochastic infection state of node ii, where Xi=1X_i = 1 indicates the presence of pathology, and let u(t)∈RNu(t) \in\mathbb{R}^{N} denote the corresponding neuronal activity state.

Conditional on the current infection and activity states, a susceptible node ii becomes infected at rate

λi(X,u)=β∑j=1NWijΦ(uj)Xj,(1)\lambda_i(X,u) = \beta\sum_{j=1}^{N} W_{ij}\Phi(u_j)X_j, \tag*{(1)}

whereas an infected node clears pathology at rate ζ\zeta. Here, W≥0W \ge0 is an irreducible weighted adjacency matrix, β>0\beta> 0 is the baseline transmission rate, and ζ>0\zeta> 0 is the clearance rate. The smooth function Φ:R→(0,∞)\Phi: \mathbb{R} \to(0,\infty) is applied nodewise and converts source-node activity into an outward transmission multiplier. It may be linear, polynomial, sigmoidal, or any other positive smooth response in the region of state space under consideration. Earlier activity–spreading models used a linear dependence of protein transport on neuronal activity [16, 18], with Φ(uj)=1+δuj\Phi(u_j) = 1 + \delta u_j. We use the same symbol Φ\Phi for its componentwise extension to vectors, so that [Φ(u)]j=Φ(uj)[\Phi(u)]_j = \Phi(u_j) for u∈RNu \in\mathbb{R}^{N} and j=1,…,Nj = 1,\ldots,N. Whether Φ\Phi denotes the scalar function or its vector extension is therefore clear from its argument.

The neuronal state evolves according to

u˙=G(u,X),(2)\dot{u} = \mathcal{G}(u,X), \tag*{(2)}

where G:RN×RN→RN\mathcal{G} : \mathbb{R}^{N} \times\mathbb{R}^{N} \to\mathbb{R}^{N} is sufficiently smooth. The infection component is therefore a continuous-time Markov chain whose transition rates depend on the continuous activity state. The combined process (X,u)(X,u) is a piecewise-deterministic Markov process: infection changes through discrete jumps, while activity evolves continuously between jumps. To make the calculations more tractable, we introduce simplifying assumptions and study the resulting approximate deterministic system of ordinary differential equations.

Let xi(t)=E[Xi(t)]x_i(t) = \mathbb{E}[X_i(t)] denote the marginal infection probability. Applying a node-level mean-field closure to the first-moment equations and replacing the stochastic activity response by its deterministic mean-field approximation gives the coupled system

x˙=β(1−x)⊙Wdiag⁡(Φ(u))x−ζx,u˙=G(u,x).(3)\begin{aligned} \dot{x} &= \beta(1-x) \odot W \operatorname{diag}(\Phi(u))x - \zeta x,\\ \dot{u} &= \mathcal{G}(u,x). \tag*{(3)} \end{aligned}

System (3) is a node-level mean-field approximation that neglects correlations between the infection states of different nodes and replaces the stochastic input to activity by its mean. It is expected to be accurate when these correlations and finite-size fluctuations are weak, as in large, sufficiently well-mixed networks. It is expected to be less accurate in small or highly structured networks, where higher-order statistics more strongly influence system dynamics.

The linear epidemic threshold

We study the dynamics near the epidemic threshold of the node-level mean-field SIS-activity system given by (3). The epidemic threshold refers to the transcritical bifurcation observed in moment closures of SIS spreading models, where a disease-free state, xi=0x_i = 0 for all nodes 1≤i≤N1 \le i \le N, and an endemic state, xi≠0x_i \ne0 for some node ii, switch stability. To determine the epidemic threshold, we linearize the coupled system about a disease-free equilibrium. Suppose that the activity subsystem (2) has a hyperbolically stable equilibrium a∈RNa \in\mathbb{R}^{N} in the absence of pathology, so that G(a,0)=0\mathcal{G}(a,0) = 0 and the activity Jacobian JA=DuG(a,0)J_A = D_u\mathcal{G}(a,0) satisfies Re⁡λ<0\operatorname{Re}\lambda< 0 for every λ∈spec⁡(JA)\lambda\in\operatorname{spec}(J_A). The coupled system then has the disease-free equilibrium (x,u)=(0,a)(x,u) = (0,a). We shift the activity variable by defining y=u−ay = u-a, which maps this equilibrium to (x,y)=(0,0)(x,y) = (0,0).

To write the linearization compactly, define A=diag⁡(Φ(a))A = \operatorname{diag}(\Phi(a)), which contains the transmission multipliers at the disease-free activity state, and C=DxG(a,0)C = D_x\mathcal{G}(a,0), which describes the linear response of activity to spreading. In the shifted coordinates, the Jacobian is

JDF(β)=(βWA−ζI0CJA).(4)J_{\mathrm{DF}}(\beta) = \begin{pmatrix} \beta W A - \zeta I & 0\\ C & J_A \end{pmatrix}. \tag*{(4)}

The upper-right block vanishes because activity modifies the transmission of existing pathology but cannot generate pathology when x=0x = 0. Since Eq. (4) is block triangular,

spec⁡(JDF(β))=spec⁡(βWA−ζI)∪spec⁡(JA).\operatorname{spec}(J_{\mathrm{DF}}(\beta)) = \operatorname{spec}(\beta W A - \zeta I) \cup\operatorname{spec}(J_A).

The pathology-to-activity coupling CC therefore does not change the eigenvalues controlling the loss of stability of the disease-free state. Let ρ0=ρ(WA)\rho_0 = \rho(W A), and let r0r_0 and ℓ0\ell_0 be the corresponding right and left Perron eigenvectors:

WAr0=ρ0r0,ℓ0TWA=ρ0ℓ0T,ℓ0Tr0=1.(5)W A r_0 = \rho_0 r_0,\qquad\ell_0^{\mathsf{T}} W A = \rho_0 \ell_0^{\mathsf{T}},\qquad\ell_0^{\mathsf{T}} r_0 = 1. \tag*{(5)}

The leading infection eigenvalue is λ0(β)=βρ0−ζ\lambda_0(\beta) = \beta\rho_0 - \zeta. Consequently, the disease-free equilibrium loses stability at

βc=ζρ0.(6)\beta_c = \frac{\zeta}{\rho_0}. \tag*{(6)}

Note that no particular form of the activity dynamics is required: its effect on the threshold enters only through the disease-free activity pattern aa, as contained in AA [18]. The quantities JAJ_A and CC instead determine the activity response needed for the nonlinear reduction below.

Activity response along the critical invasion mode

To construct the center-manifold reduction near the threshold, we need the right nullspace of the full Jacobian JDF(βc)J_{\mathrm{DF}}(\beta_c). It is the center eigenspace tangent to the center manifold, and its activity component describes how activity changes during invasion.

Because WW is irreducible and AA has strictly positive diagonal entries, WAW A is irreducible and its Perron eigenvalue ρ0\rho_0 is simple. For the infection block Lc=βcWA−ζIL_c = \beta_c W A - \zeta I, the identity βcρ0=ζ\beta_c \rho_0 = \zeta therefore gives ker⁡Lc=span⁡{r0}\ker L_c = \operatorname{span}\{r_0\}, and so the right Perron eigenvector r0r_0 is the infection component of the center eigenspace. Since JAJ_A is invertible, the block structure in (4) implies that ker⁡JDF(βc)\ker J_{\mathrm{DF}}(\beta_c) is also one-dimensional. Choose its right nullvector as vc=(r0,h)Tv_c = (r_0, h)^{\mathsf{T}}. The condition JDF(βc)vc=0J_{\mathrm{DF}}(\beta_c)v_c = 0 gives Lcr0=0L_c r_0 = 0 in the infection block and JAh+Cr0=0J_A h + C r_0 = 0 in the activity block, and hence

h=−JA−1Cr0.(7)h = -J_A^{-1} C r_0. \tag*{(7)}

Thus, hh is the leading activity response to invasion in the critical infection pattern r0r_0. Accordingly, hi>0h_i > 0 means that activity at node ii increases as pathology invades, whereas hi<0h_i < 0 means that it decreases, to leading order; hi=0h_i = 0 means that there is no leading-order change. The vector Cr0C r_0 is the immediate activity perturbation induced by this pattern, while the operator −JA−1-J_A^{-1} accumulates this forcing over the subsequent activity relaxation.

Center-manifold reduction near the threshold

Having identified the critical direction in the full state space, we now reduce the coupled dynamics along it. The resulting scalar equation describes how the endemic branch emerges from the threshold and provides the coefficients used to classify the invasion regimes below.

Let μ=β−βc\mu= \beta- \beta_c denote the distance from the epidemic threshold. At μ=0\mu= 0, the center manifold through the disease-free equilibrium is tangent to the critical direction (r0,h)T(r_0, h)^{\mathsf{T}}. For β\beta near βc\beta_c, we use z=ℓ0Txcmz = \ell_0^{\mathsf{T}} x_{\mathrm{cm}} as its scalar coordinate, so states on the manifold satisfy

xcm(z,μ)=zr0+O(z2,μz),x_{\mathrm{cm}}(z,\mu) = z r_0 + O(z^2,\mu z),
ycm(z,μ)=zh+O(z2,μz).(8)y_{\mathrm{cm}}(z,\mu) = z h + O(z^2,\mu z). \tag*{(8)}

Nearby states approach this form as their stable components decay. The reduced amplitude equation has the form

z˙=ρ0μz+c2z2+c3z3+O(μ2z,μz2,z4).(9)\dot{z} = \rho_0 \mu z + c_2 z^2 + c_3 z^3 + O(\mu^2 z, \mu z^2, z^4). \tag*{(9)}

We use this cubic truncation as the starting point to study the dynamics of the coupled system given by Eq. (3) near the epidemic threshold. The coefficient ρ0\rho_0 of μz\mu z is always positive. By contrast, c2c_2 and c3c_3 may either be positive, negative, or zero. The coefficient c2c_2 determines the generic transcritical orientation, whereas c3c_3 controls the first curvature and saturation of the endemic branches. Their dependence on the feedback is described in the next two subsections, with the reduction summarized in Appendix A.2.

Quadratic coefficient

The quadratic coefficient in (9) is

c2=−ζℓ0T(r0⊙r0)+βcℓ0TWD1[h]r0.(10)c_2 = -\zeta\ell_0^{\mathsf{T}}(\mathbf{r}_0 \odot\mathbf{r}_0) + \beta_c \ell_0^{\mathsf{T}} W D_1[h]\mathbf{r}_0. \tag*{(10)}

Here, D1[h]=diag⁡(Φ′(a)⊙h)D_1[h] = \operatorname{diag}(\Phi'(a) \odot h) denotes the first-order change in the transmission multipliers along the activity response hh. The calculation is given in Appendix A.2.

The quadratic coefficient c2c_2 determines whether the transcritical bifurcation at μ=0\mu= 0 is supercritical (c2<0c_2 < 0), subcritical (c2>0c_2 > 0), or degenerate (c2=0c_2 = 0). The first term represents the intrinsic SIS saturation. It is negative because susceptible depletion suppresses further infection, making the transcritical bifurcation supercritical in the absence of the feedback loop. The second term represents the pathology–activity feedback contribution. The sign of this contribution is controlled by Φ′(a)⊙h\Phi'(a) \odot h, where hh encodes how the activity dynamics respond to pathology, while Φ′(a)\Phi'(a) encodes how activity changes transmission. When these couplings have the same sign componentwise, the feedback is reinforcing and the second term is positive. In this case, the feedback offsets the intrinsic SIS saturation and may drive c2c_2 through zero, producing the backward orientation required for a finite-amplitude invasion boundary. When their signs are opposite componentwise, the feedback opposes spreading: the second term is negative and makes c2c_2 more negative. If the signs vary across nodes, the overall sign is determined by the Perron-weighted sum in (10). To find out when, if at all, the endemic branch folds to generate multistability, we calculate the cubic coefficient.

Cubic coefficient

The cubic coefficient in (9) is

n3=−βc[r0⊙WAξ+ξ⊙WAr0]+βcW[D1[h]ξ+D1[ψ]r0+12D2[h,h]r0]−βcr0⊙WD1[h]r0,c3=ℓ0Tn3.(11)\begin{aligned} n_3 ={}& -\beta_c\left[\mathbf{r}_0 \odot W A \boldsymbol{\xi} + \boldsymbol{\xi} \odot W A\mathbf{r}_0\right] \\ &+ \beta_c W\left[D_1[h]\boldsymbol{\xi} + D_1[\psi]\mathbf{r}_0 + \frac{1}{2}D_2[h,h]\mathbf{r}_0\right] \\ &- \beta_c \mathbf{r}_0 \odot W D_1[h]\mathbf{r}_0, \\ c_3 ={}& \ell_0^{\mathsf{T}} n_3. \tag*{(11)} \end{aligned}

Here, at μ=0\mu= 0, we write xcm=zr0+z2ξ+O(z3)x_{\mathrm{cm}} = z\mathbf{r}_0 + z^2\boldsymbol{\xi} + O(z^3) and ycm=zh+z2ψ+O(z3)y_{\mathrm{cm}} = zh + z^2\psi+ O(z^3), so that ξ\boldsymbol{\xi} and ψ\psi are the second-order corrections to the infection pattern and the activity response, respectively. We further define D2[h,h]=diag⁡(Φ′′(a)⊙h⊙h)D_2[h,h] = \operatorname{diag}(\Phi''(a) \odot h \odot h). The calculation is given in Appendix A.2.

The cubic coefficient gives the first indication of whether the endemic branch can fold. Unlike c2c_2, the sign of c3c_3 cannot generally be inferred from the coupling directions. It combines susceptible depletion acting on the curved infection pattern, activity feedback through ξ\xi and ψ\psi, the quadratic part of G\mathcal{G}, and the transmission curvature Φ′′(a)\Phi^{\prime\prime}(a). A negative c3c_3 provides cubic saturation, whereas a positive c3c_3 reinforces growth at cubic order. Its effect on the endemic branches depends on the sign of c2c_2.

We next determine when c2c_2 and c3c_3 produce a fold of the endemic branch in the physical region.

Local invasion regimes near the epidemic threshold

A fold of the endemic branch creates a saddle-node bifurcation and can thereby generate multistability. Retaining terms through cubic order gives

z˙=z(ρ0μ+c2z+c3z2).(12)\dot{z}=z\left(\rho_0\mu+c_2z+c_3z^2\right). \tag*{(12)}

For z≠0z\ne0, equilibria satisfy ρ0μ+c2z+c3z2=0\rho_0\mu+c_2z+c_3z^2=0. At a saddle-node, two equilibria coalesce, so c2+2c3z=0c_2+2c_3z=0 as well. Hence,

zSN=−c22c3,μSN=c224ρ0c3.(13)z_{\mathrm{SN}}=-\frac{c_2}{2c_3},\qquad\mu_{\mathrm{SN}}=\frac{c_2^2}{4\rho_0c_3}. \tag*{(13)}

Only zSN>0z_{\mathrm{SN}}>0 corresponds to nonnegative infection probabilities. Thus, the cubic center-manifold truncation produces a fold in the physical region only when c2c_2 and c3c_3 have opposite signs.

Backward transition and finite-amplitude invasion. A subcritical transcritical bifurcation has c2>0c_2>0. Its fold lies in the physical region when c3<0c_3<0. The fold then lies below the threshold, with μSN<0\mu_{\mathrm{SN}}<0. The unstable positive branch from the backward transcritical bifurcation turns at the saddle-node and continues as a stable endemic branch. For μSN<μ<0\mu_{\mathrm{SN}}<\mu<0, the disease-free state and a larger endemic state are stable, while a smaller unstable endemic state separates their basins. Thus, increasing β\beta preserves the disease-free state until βc\beta_c, whereas decreasing it preserves the endemic state down to βSN=βc+μSN\beta_{\mathrm{SN}}=\beta_c+\mu_{\mathrm{SN}}. This gives finite-amplitude invasion and hysteresis below the threshold.

Forward transition and endemic–endemic multistability. A supercritical transcritical bifurcation has c2<0c_2<0. Its fold lies in the physical region when c3>0c_3>0. The fold then lies above the threshold, with μSN>0\mu_{\mathrm{SN}}>0. The stable positive branch turns at the saddle-node and continues as an unstable branch. For 0<μ<μSN0<\mu<\mu_{\mathrm{SN}}, the cubic truncation therefore contains a smaller stable endemic equilibrium and a larger unstable one. This first fold does not yet give bistability because the disease-free state is unstable. Endemic–endemic multistability requires a second fold, whose local existence depends on the fourth-order coefficient c4c_4. Rather than carrying out that higher-order calculation here, we verify the second fold numerically in the next subsection.

Degenerate transition. The boundary case c2=0c_2=0 occurs when activity feedback exactly balances the intrinsic SIS saturation and separates the supercritical and subcritical orientations. It is nongeneric, and c3c_3 or higher-order terms determine the leading branch geometry.

Numerical validation of the invasion regimes

We verify in the full deterministic network equations that the physical folds predicted above can produce finite-amplitude invasion below the threshold and endemic–endemic bistability above it, rather than occurring only in the truncated center-manifold equation. We use an Erdős–Rényi graph with 20 nodes, where each edge is included with probability 0.2. Its adjacency matrix is normalized so that ρ(W)=1\rho(W) = 1, and we set the SIS clearance ζ=1\zeta= 1. The activity dynamics are kept minimal with a linear spreading-to-activity coupling, u˙i=−ui+xi\dot{u}_i = -u_i + x_i, and we vary only the activity-to-spreading coupling Φ\Phi.

The linear coupling Φ(u)=1+u/2\Phi(u) = 1 + u/2 produces monostable forward invasion, while the quadratic polynomials Φ(u)=(1+u)2\Phi(u) = (1 + u)^2 and Φ(u)=1+3u/5+5u2/2\Phi(u) = 1 + 3u/5 + 5u^2/2 produce finite-amplitude invasion below the epidemic threshold β<βc\beta< \beta_c and endemic–endemic bistability above the threshold β>βc\beta> \beta_c. For each choice of Φ\Phi, we compute the numerical bifurcation diagram of the full 40-dimensional system as β\beta is varied. Each endemic equilibrium is represented by its infection amplitude z=ℓ0Txz = \ell_0^{\mathsf{T}} x. The numerical continuation and stability calculation is described in Appendix A.3. Figure 1 plots the resulting branches together with the local center-manifold prediction. The dash-dotted lines mark the values of β\beta used for direct integration of the full system in panels (d–f). The center-manifold reduction approximates the branches well close to the bifurcation.

Figure 1. Numerical bifurcations on an Erdős–Rényi graph G(20,0.2)G(20, 0.2). (a–c) Endemic branches for Φ(u)=1+u/2\Phi(u) = 1 + u/2, Φ(u)=(1+u)2\Phi(u) = (1 + u)^2, and Φ(u)=1+3u/5+5u2/2\Phi(u) = 1 + 3u/5 + 5u^2/2, respectively. Black dashed curves show the center-manifold predictions. Solid blue and dashed red branches are stable and unstable, black dotted vertical lines mark βc=1\beta_c = 1, and dash-dotted purple lines mark the values of β\beta used in (d–f). (d–f) Mean infection from lower and higher initial conditions at β=1.08\beta= 1.08, 0.900.90, and 1.021.02, respectively. In (d,e), the uniform initial states have xi(0)=ui(0)=0.01x_i(0) = u_i(0) = 0.01 or 0.550.55; in (f), they have xi(0)=ui(0)=0.02x_i(0) = u_i(0) = 0.02 or 0.400.40.

We then integrate the full equations through time from lower and higher initial conditions. The trajectories confirm the stability and initial-condition dependence of the branches in Figure 1.

Endemic multistability on regular graphs

The center-manifold reduction shows how endemic branches emerge near the epidemic threshold, but it cannot determine how many endemic states exist in the full nonlinear system or what happens farther from the threshold. To address these questions in a simpler setting, we now restrict the system to homogeneous dynamics on a regular graph. For general smooth coupling functions, we show that the number of endemic equilibria is even below the epidemic threshold and odd above it, and we exclude periodic behavior. We further show that multistability is impossible for opposite-sign monotone coupling. For polynomial couplings, we find a bound on the number of endemic equilibria, and for sigmoidal couplings, we identify bistable regimes.

Homogeneous mean-field reduction

Suppose that WW represents a regular graph in which every node has weighted degree ρ\rho, so that W1=ρ1W\mathbf{1} = \rho\mathbf{1}. We assume that all edge weights are nonnegative and that the graph is connected, or strongly connected if its edges are directed. Equivalently, WW is an irreducible nonnegative matrix. The Perron–Frobenius theorem then implies that ρ\rho is the spectral radius of WW. A homogeneous initial condition remains homogeneous: if every node starts with the same infection probability xx and activity deviation yy, all nodes continue to have the same values. We therefore restrict the following analysis to these homogeneous dynamics, which are described by one infection variable and one activity variable.

As before, Φ\Phi denotes the activity-to-transmission coupling. We write H(x)H(x) for the pathology-to-activity coupling and require that H(0)=0H(0) = 0. The homogeneous dynamics are

x˙=x[βρ(1−x)Φ(a+y)−ζ],y˙=−y+H(x).(14)\begin{aligned} \dot{x} &= x\left[\beta\rho(1-x)\Phi(a+y)-\zeta\right],\\ \dot{y} &= -y+H(x). \tag*{(14)} \end{aligned}

Because H(0)=0H(0) = 0, the disease-free equilibrium is (x,y)=(0,0)(x,y) = (0,0), with invasion threshold βc=ζρΦ(a)\beta_c = \frac{\zeta}{\rho\Phi(a)}.

General constraints on the homogeneous dynamics

The homogeneous system constrains the number and stability of its endemic equilibria and rules out sustained oscillations. Let α=ζβρ\alpha= \frac{\zeta}{\beta\rho}. At an endemic equilibrium of the homogeneous system (14), the activity nullcline gives y=H(x)y = H(x), so the infection coordinate is a root in (0,1)(0,1) of

F(x):=(1−x)Φ(a+H(x))−α.(15)F(x) := (1-x)\Phi(a+H(x))-\alpha. \tag*{(15)}

The endpoint signs and the direction of each crossing determine how many equilibria can occur on either side of the threshold and which of them are stable. Because F(0)=Φ(a)−αF(0) = \Phi(a)-\alpha is positive above βc\beta_c and negative below it, while F(1)=−α<0F(1) = -\alpha< 0, the number of simple endemic roots is odd above βc\beta_c and even below it. At a simple endemic root x∗x_*, the trace and determinant of the Jacobian of the homogeneous system in (14) satisfy

tr⁡J=−βρx∗Φ(a+H(x∗))−1<0,det⁡J=−βρx∗F′(x∗).(16)\operatorname{tr} J = -\beta\rho x_*\Phi(a+H(x_*))-1 < 0,\qquad\det J = -\beta\rho x_*F'(x_*). \tag*{(16)}

Hence downward crossings of FF are stable and upward crossings are saddles. Simple endemic equilibria alternate in stability, and the endemic equilibrium with the highest pathology level is always stable.

The homogeneous system also cannot support sustained oscillations. Choosing the Dulac function B(x,y)=1/xB(x,y)=1/x on x>0x>0, the Bendixson–Dulac criterion excludes periodic orbits in the physical region using the assumed positivity of Φ\Phi.

The equilibrium function also gives a global sign rule for monotone couplings. Differentiating (15) gives

F′(x)=−Φ(a+H(x))+(1−x)Φ′(a+H(x))H′(x).(17)F'(x)=-\Phi(a+H(x))+(1-x)\Phi'(a+H(x))H'(x). \tag*{(17)}

Thus, FF turns upwards precisely when

(1−x)Φ′(a+H(x))H′(x)>Φ(a+H(x)).(18)(1-x)\Phi'(a+H(x))H'(x)>\Phi(a+H(x)). \tag*{(18)}

If one coupling is non-decreasing and the other is non-increasing, then Φ′(a+H(x))H′(x)≤0\Phi'(a+H(x))H'(x)\le0, so the inequality cannot hold. Hence, FF is strictly decreasing, giving no endemic equilibrium at or below βc\beta_c and exactly one stable endemic equilibrium above it. If the couplings are monotone in the same direction, the inequality may hold when the feedback is sufficiently strong. Same-sign coupling therefore permits multistability, but does not guarantee it. For opposite-sign coupling, however, multistability is ruled out.

The Bendixson–Dulac calculation is given in Appendix B.1.

Polynomial bounds on endemic multistability

Polynomial coupling functions provide a natural starting point for studying nonlinear feedback and give an explicit bound on the number of equilibria. If Φ\Phi and HH have degrees mm and nn, respectively, then the equilibrium function FF in (15) has degree at most mn+1mn+1. Hence

Nendemic≤mn+1.(19)N_{\mathrm{endemic}}\le mn+1. \tag*{(19)}

If both couplings are linear, then m=n=1m=n=1 and the bound allows at most two endemic equilibria. However, we know that the number of endemic equilibria is odd above the threshold, so there is exactly one stable endemic equilibrium. Below it, we have an even number of endemic equilibria, so there are either no endemic equilibria or a saddle–stable pair. Therefore, reinforcing linear coupling can produce disease-free–endemic bistability, but not endemic–endemic bistability. Higher polynomial degrees, however, permit additional endemic states.

We next show that higher-degree polynomials can produce more stable endemic equilibria. We fix the linear activity dynamics y˙=−y+x\dot{y}=-y+x, set a=0a=0 and ρ=ζ=1\rho=\zeta=1, and vary only Φ\Phi. The transmission functions and their exact factorizations are given in Appendix B.2.

Figure 2 shows the phase planes for quadratic, cubic, and quartic transmission functions. They produce two stable endemic states above the threshold, the disease-free state together with two stable endemic states below the threshold, and three stable endemic states above the threshold, respectively. Direct integrations in panels (d–f) confirm that the initial condition selects among these attractors.

Figure 2. Multistability generated by polynomial activity-to-transmission coupling with the linear activity dynamics y˙=−y+x\dot{y}=-y+x. (a–c) Phase planes for quadratic, cubic, and quartic transmission functions. (d–f) The corresponding direct integrations from different homogeneous initial conditions with y(0)=x(0)y(0)=x(0). Dotted horizontal lines mark the stable equilibria. Filled disease-free diamonds are stable and open diamonds are unstable.

Multistability with sigmoidal couplings

Polynomials give a general description of nonlinear feedback, but biological effects, including those in neurodegeneration, often saturate. The global sign rule above already shows that opposite-sign monotone couplings, such as sigmoids, cannot produce multistability. We therefore test whether same-sign sigmoidal couplings can produce multistability and, if so, how many stable equilibria they can support.

Let SxS_x and SuS_u be increasing logistic functions normalized between zero and one. We take

Φ(u)=1+δSu(u),H(x)=ηSx(x),\Phi(u)=1+\delta S_u(u), \qquad H(x)=\eta S_x(x),

where the signed gains δ\delta and η\eta set the strength and direction of the activity-to-spreading and spreading-to-activity couplings.

Figure 3 compares the four sign combinations. Opposite-sign couplings give one stable endemic equilibrium. For same-sign couplings, gains of magnitude 0.30 also give one endemic equilibrium, whereas increasing the magnitude of either gain to 0.90 produces two stable endemic equilibria separated by a saddle. Thus, same-sign sigmoidal coupling can produce multistability when the feedback is sufficiently strong.

Figure 3. Nullclines and endemic equilibria for sigmoidal coupling. (a,e) Opposite signs with (δ,η)=(0.90,−0.90)(\delta,\eta)=(0.90,-0.90) and (−0.90,0.90)(-0.90,0.90). (b,f) Same signs with (δ,η)=(0.30,0.30)(\delta,\eta)=(0.30,0.30) and (−0.30,−0.30)(-0.30,-0.30). (c,g) Increasing ∣δ∣|\delta| to 0.90 produces multistability; (d,h) increasing ∣η∣|\eta| does the same. Blue and orange curves show the infection and activity nullclines. Filled circles are stable endemic equilibria, crosses are saddles, and open diamonds are unstable disease-free equilibria. All panels use β/βc=1.15\beta/\beta_c=1.15; the two sigmoids have steepnesses 24 and 6 and transition points x=0.25x=0.25 and u=0.40u=0.40, respectively.

Invasion and multistability in QIF neuronal models

We next test the predicted invasion regimes in QIF neuronal models at two spatial scales. First, we use a regular network of mean-field QIF populations, where each node represents a neuronal population, to calculate the bifurcation structure. We then model a neural circuit in which each node is an individual spiking QIF neuron coupled to stochastic SIS spreading.

Bifurcation structure in mean-field QIF populations

We model a brain network in which each node represents the population-averaged dynamics of many neurons, using the exact mean-field equations for a heterogeneous QIF population [31]. The populations are coupled both to spreading and to each other. Along each neural edge j→ij \to i, all neurons in population jj project uniformly to all neurons in population ii, giving the common input KWijrjK W_{ij}r_j while preserving the exact QIF reduction [32]. As before, H(x)H(x) describes infection-induced changes in excitability and Φ(r)\Phi(r) the effect of firing on transmission.

x˙i=β(1−xi)∑jWijΦ(rj)xj−xi,εr˙i=Δπ+2rivi,εv˙i=vi2+ηˉ+H(xi)+Jlocri+K∑jWijrj−(πri)2.(20)\begin{aligned} \dot{x}_i &= \beta(1-x_i)\sum_j W_{ij}\Phi(r_j)x_j-x_i,\\ \varepsilon\dot{r}_i &= \frac{\Delta}{\pi}+2r_iv_i,\\ \varepsilon\dot{v}_i &= v_i^2+\bar{\eta}+H(x_i)+J_{\mathrm{loc}}r_i+K\sum_j W_{ij}r_j-(\pi r_i)^2. \tag*{(20)} \end{aligned}

Here xix_i is the infection probability, while rir_i and viv_i are the firing rate and mean membrane potential of population ii. The parameter ε\varepsilon is the ratio between neuronal and spreading timescales, Δ\Delta is the half-width of the Lorentzian excitability distribution, and ηˉ\bar{\eta} is its mean. The gains JlocJ_{\mathrm{loc}} and KK set the neuronal coupling within and between populations, respectively. Time is measured in units of the mean recovery time, so the SIS recovery rate is one.

To calculate the bifurcation structure, we restrict the analysis to homogeneous initial conditions on a regular graph. All nodes then remain identical, so their equilibrium branches can be followed in a homogeneous system. We use a normalized 40-node ring network with four connections of weight 1/41/4 per node. On the homogeneous system, the interpopulation input reduces to KrKr. Both couplings are bounded sigmoids with steepness 20. Let r0r_0 denote the stationary firing rate without infection. The pathology sigmoid SxS_x is centered at x=0.25x = 0.25 and is therefore weak when pathology is rare, while the firing sigmoid SrS_r is centered at r=r0r = r_0 and detects changes from baseline firing. Both vanish at zero and approach one. We define

H(x)=κSx(x),Φ(r)=1+δSr(r).(21)H(x) = \kappa S_x(x), \qquad\Phi(r) = 1 + \delta S_r(r). \tag*{(21)}

Here, κ\kappa and δ\delta are the spreading-to-activity and activity-to-spreading gains, respectively. Their signs set the two coupling directions, and Φ(0)=1\Phi(0) = 1. Consequently, the epidemic threshold is βc=1/Φ(r0)\beta_c = 1/\Phi(r_0).

For Jloc+K≤0J_{\mathrm{loc}} + K \le0, the homogeneous QIF system has a unique positive firing rate r∗(x)r_*(x) for every fixed infection level (Appendix C). The sign rule in Sec. 3 permits multistability when the feedback between the neural dynamics and spreading is reinforcing and sufficiently strong. Since the firing rate is unique at each infection level, this multistability must come from the feedback loop. We vary κ\kappa and δ\delta to test for finite-amplitude invasion and endemic–endemic bistability.

On the homogeneous ring, the equilibrium firing rate at infection level xx is r∗(x)r_*(x). Since the recovery rate and spectral radius are both one, the endemic equilibria are the roots of

F(x;β)=(1−x)Φ(r∗(x))−1β.(22)F(x;\beta) = (1-x)\Phi(r_*(x)) - \frac{1}{\beta}. \tag*{(22)}

We set Jloc=−2J_{\mathrm{loc}} = -2, K=1K = 1, ηˉ=0.25\bar{\eta} = 0.25, Δ=0.1\Delta= 0.1, and ε=0.05\varepsilon= 0.05. These values keep the homogeneous QIF system monostable, giving the disease-free firing rate r0≃0.121r_0 \simeq0.121. We then vary κ\kappa and δ\delta to obtain the equilibrium branches in Figure 4(a–d).

Figure 4. Effect of the coupling gains in the mean-field QIF populations. (a,b) Equilibrium branches as κ\kappa or δ\delta changes sign with the other gain fixed and positive. (c,d) Corresponding sweeps with the other gain fixed and negative. Circles mark saddle-node bifurcations and dotted lines mark the epidemic threshold. (e) Finite-amplitude invasion below the threshold. (f) Endemic–endemic bistability above it. Simulations use (Jloc,K,κ,δ)=(−2,1,1.50,0.95)(J_{\mathrm{loc}}, K, \kappa, \delta) = (-2, 1, 1.50, 0.95) on a 40-node ring, with mean initial infection levels 0.10 and 0.30. Curves show the network mean and shading shows the range over nodes.

Panels (a,c) vary κ\kappa at fixed δ=±0.95\delta= \pm0.95, while panels (b,d) vary δ\delta at fixed κ=±1.50\kappa= \pm1.50. Opposite-sign couplings give a single endemic branch, whereas sufficiently strong same-sign couplings produce two saddle-node bifurcations. For (κ,δ)=(1.50,0.95)(\kappa,\delta) = (1.50, 0.95), these occur at β/βc≃0.991\beta/\beta_c \simeq0.991 and 1.050. Direct simulations with the same parameters verify finite-amplitude invasion below the threshold in panel (e) and endemic–endemic bistability above it in panel (f).

Multistability in stochastic spiking networks

We now model a stochastic neural circuit in which each node is an individual spiking QIF neuron coupled to stochastic SIS spreading. The QIF neuron is equivalent to the theta-neuron model under a change of variables. In this form, the membrane state of neuron ii is represented by a phase θi\theta_i that moves around a circle, and a crossing of θi=π\theta_i = \pi is recorded as a spike. The variable Xi(t)∈{0,1}X_i(t) \in\{0,1\} is now a susceptible/infected state rather than an infection probability. The neural dynamics obey

εθ˙i=1−cos⁡θi+(1+cos⁡θi)[ηi+κXi+J∑jWijsj+liext(t)].(23)\varepsilon\dot{\theta}_i = 1-\cos\theta_i + (1+\cos\theta_i)\left[\eta_i+\kappa X_i+J\sum_j W_{ij}s_j+l_i^{\mathrm{ext}}(t)\right]. \tag*{(23)}

Here the excitabilities ηi\eta_i are chosen to represent a Lorentzian distribution with mean ηˉ\bar{\eta} and half-width Δ\Delta, as detailed in Appendix D.1. The parameter ε\varepsilon sets the neuronal timescale, JJ is the synaptic coupling strength, and sj(t)s_j(t) is a time-dependent measure of the recent firing of neuron jj, whose evolution is defined below by (24). Infection changes the current through κXi\kappa X_i, while

Iiext(t)I_i^{\mathrm{ext}}(t) is an externally applied neuronal input. The recent spike history of neuron ii is represented by the exponential filter

τss˙i=−si+ε∑kδD(t−tik),(24)\tau_s \dot{s}_i = -s_i + \varepsilon\sum_k \delta_D\left(t - t_i^k\right), \tag*{(24)}

where τs\tau_s is the filter timescale and tikt_i^k are the spike times.

When node jj is infected, its own recent firing changes its outward transmission through the same sigmoidal form SrS_r used in (21). Conditional on the current neural state, the SIS transitions are

Xi:0⟶1at rateλi(t)=βτX∑jWij(1+δSr(sj(t)))Xj(t),(25)X_i: 0 \longrightarrow1 \quad\text{at rate} \quad\lambda_i(t)=\frac{\beta}{\tau_X}\sum_j W_{ij}\left(1+\delta S_r(s_j(t))\right)X_j(t), \tag*{(25)}
Xi:1⟶0at rateζτX.X_i: 1 \longrightarrow0 \quad\text{at rate} \quad\frac{\zeta}{\tau_X}.

Here β\beta and ζ\zeta set the transmission and recovery rates, τX\tau_X is the spreading timescale, and δ\delta is the activity-to-spreading gain. We simulate the coupled system by alternating a QIF integration step with exact Gillespie sampling of the SIS events [33, 34]. The full algorithm is given in Appendix D.

To test whether the same-sign multistability found in the mean-field QIF model (Figure 4) persists in a spiking network, we take κ>0\kappa> 0 and δ>0\delta> 0. We draw the positions of N=1000N = 1000 nodes independently from a uniform distribution over the unit square and connect every pair separated by less than 0.155. The resulting connected graph has 32 893 undirected edges and mean degree 65.8. We normalize the edge weights by setting W=A/ρ(A)W = A/\rho(A), so that ρ(W)=1\rho(W) = 1. The neuronal parameters are ηˉ=−3.25\bar{\eta} = -3.25, Δ=0.1\Delta= 0.1, J=6J = 6, and κ=3\kappa= 3, while the transmission gain is δ=12\delta= 12. The firing sigmoid is centered at s=0.1s = 0.1 and has steepness 2. Neuronal activity evolves over milliseconds to seconds, whereas pathological spreading occurs over months to years [16, 35]. We therefore set τX=100\tau_X = 100, ϵ=0.02\epsilon= 0.02, and τs=0.05\tau_s = 0.05, so that many spikes occur between infection and recovery events.

We first test whether the endemic–endemic bistability found in the mean-field QIF populations is also present in the stochastic spiking network. At β=0.90\beta= 0.90, initial infected fractions 0.25 and 0.75 remain near mean prevalences 0.175 and 0.697, respectively, showing distinct low- and high-endemic regimes (Figure 5). We then ask whether neuronal input can move the network between these regimes. Starting from the lower regime, we set Iiext(t)=1I_i^{\mathrm{ext}}(t) = 1 at every node for 500≤t<1000500 \le t < 1000 and Iiext(t)=−1I_i^{\mathrm{ext}}(t) = -1 for 3000≤t<35003000 \le t < 3500, with zero input otherwise. The first input moves the stochastic SIS process to the high-endemic regime, while the second returns it to the low-endemic regime (Figure 5). Both regimes remain after the corresponding input is removed. Thus, the stochastic spiking network exhibits endemic–endemic bistability, and transient neuronal input can move it between the two endemic regimes. A realization of the stochastic simulation with external input is shown in Supplementary Video S1.

Figure 5. Endemic–endemic bistability in the stochastic SIS–QIF network. (a,b) Infection prevalence and mean filtered firing from low and high initial infection levels. (c,f) Positive (purple) and negative (blue) neuronal inputs move the network between the two endemic regimes. (d,e) Representative spatial states; blue and orange nodes are susceptible and infected, and white rings mark spikes. The network has N=1000N = 1000 nodes, connection radius 0.155, and mean degree 65.8. Parameters are β=0.90\beta= 0.90, ηˉ=−3.25\bar{\eta} = -3.25, Δ=0.1\Delta= 0.1, J=6J = 6, κ=3\kappa= 3, δ=12\delta= 12, τX=100\tau_X = 100, ε=0.02\varepsilon= 0.02, and τs=0.05\tau_s = 0.05. Initial infected fractions are 0.25 and 0.75, and the transmission sigmoid has steepness 2 and midpoint 0.1. Inputs have amplitude 1 and act over 500≤t<1000500 \le t < 1000 and 3000≤t<35003000 \le t < 3500.

To test for finite-amplitude invasion at lower transmission, we reduce β\beta from 0.90 to 0.76 in Figure 6. With identical neuronal initial conditions, a seed of 50 infected nodes reaches disease-free absorption, whereas a seed of 750 nodes reaches a high-endemic metastable regime. Thus, the small infection seed dies out, but infection persists when its initial level is sufficiently large. We then ask whether neuronal input can move the network across this invasion boundary. Starting from the 50-node seed and the same neuronal initial condition, we set Iiext=1I_i^{\mathrm{ext}} = 1 at every node for 500≤t<1000500 \le t < 1000 and Iiext=−1I_i^{\mathrm{ext}} = -1 for 3000≤t<35003000 \le t < 3500, with zero input otherwise. The first input moves the stochastic SIS process to the high-endemic regime, while the second returns it to the low-infection state (Figure 6). Both states remain after the corresponding input is removed. Thus, the spiking neural network exhibits finite-amplitude invasion, and transient neuronal input can move it across the invasion boundary. A realization of the stochastic simulation with external input is shown in Supplementary Video S2.

Figure 6. Finite-amplitude invasion in the stochastic SIS–QIF network at β=0.76\beta= 0.76. (a,b) A 50-node infection seed becomes extinct, whereas a 750-node seed reaches a high-endemic regime. (c,f) Starting from the same 50-node seed, positive (purple) and negative (blue) neuronal inputs move the network to the high-endemic regime and back to low infection and activity. (d,e) Representative disease-free and high-endemic spatial states. The network, remaining parameters, and input intervals are as in Figure 5.

To distinguish this behavior from a long transient, we repeat the input experiment in Appendix D.2 after reversing the sign of κ\kappa, so that the two coupling directions have opposite signs. Even with a five-times stronger input, infection begins to decay once the input is removed and eventually reaches absorption (Figure 7). This comparison therefore supports interpreting the high-endemic state in Figure 6 as the result of crossing a finite-amplitude invasion boundary.

Figure 7. Opposite-sign control at β=0.76\beta= 0.76. (a) Infection decays after the neuronal input is removed. (b) Mean filtered firing returns to baseline. The graph, 50-node infection seed, and remaining parameters are as in Figure 6, except κ=−3\kappa= -3. The input has amplitude 5 and acts over 0≤t<5000 \le t < 500.

Discussion

We have studied a spreading process which co-evolves with a general activity process on the same network. The feedback loop between activity and spreading determines whether invasion is continuous, requires a finite perturbation, or leads to several endemic states. These dynamics can occur even when the activity subsystem has a unique stable state, in which case the multistability is generated by the feedback loop itself. Numerical simulations of biophysical neuronal models recover the same dynamical behavior, both in mean-field QIF populations and in networks of individual spiking QIF neurons coupled to stochastic SIS spreading.

Activity-driven networks and other routes to multistability

Activity-driven networks provide the closest comparison to the present work. In the original framework, nodes create temporary contacts at heterogeneous rates, allowing contagion to spread over a time-varying network [36]. Subsequent work established how activity heterogeneity, temporal connectivity, and attractiveness shape epidemic thresholds and endemic prevalence [37, 38, 39], and developed strategies for controlling the coupled network and spreading dynamics [40]. Adaptive extensions allow infection, risk perception, quarantine, or preventive behavior to change node activity and contact formation [41, 42, 43]. Li et al. allowed prevalence to alter activity in a higher-order activity-driven network and found two bistable regimes [5]. In these models, activity changes the temporal network, whereas in our model it changes transmission over a fixed network. Multistability also arises through mechanisms unrelated to activity-driven contact formation. Susceptible nodes may rewire away from infected neighbours [1], recovery may depend on resources generated by the healthy population [6], and group transmission may reinforce contagion through higher-order interactions [25]. These studies show that multistability can arise when network adaptation, recovery feedback, or higher-order interactions make transmission depend nonlinearly on the spreading state. Our model provides a related mechanism in which transmission is modulated indirectly through feedback with a coevolving dynamical state on the same network.

Implications for neurodegenerative disease

The main motivation for studying this coupled system is neurodegenerative disease. Increased neuronal activity can promote tau release and propagation in Alzheimer’s disease [10, 11] and can modulate α\alpha-synuclein spreading [44] in Parkinson’s disease, providing examples of a positive activity-to-spreading coupling. Evidence that greater neuronal activity instead suppresses pathological protein spreading, which would motivate a negative activity-to-spreading coupling, is less established. By contrast, either sign appears plausible for the spreading-to-activity coupling. Amyloid-β\beta has been associated with neuronal hyperactivity [12, 45], whereas tau can impair neuronal circuits and suppress activity [46, 47]. The response to amyloid-β\beta may also be nonmonotone: low picomolar concentrations potentiate synaptic transmission, whereas higher nanomolar concentrations suppress it [48, 49]. An inverted-U, quadratic pathology-to-activity coupling may therefore be biologically relevant. Combined with a positive activity-to-spreading coupling, its rising branch reinforces spreading at low pathology, whereas its falling branch opposes spreading at higher pathology. Our theory predicts that sufficiently strong reinforcement at low pathology can create a finite-amplitude invasion boundary and disease-free–endemic bistability, while the later reversal limits further reinforcement. Thus, the spreading-to-activity coupling may depend on the protein species, disease stage, and amount of pathology.

Amyloid-associated hyperactivity and activity-dependent tau spreading suggest a specific mechanism: amyloid accumulation may first raise neuronal activity, which could then facilitate the later spread of tau. Longitudinal human studies indicate that amyloid-β\beta accumulation precedes the widespread neocortical accumulation of tau, and that an earlier rise in amyloid is associated with subsequent tau accumulation [50, 51, 52]. Amyloid-associated hyperactivity may therefore be present before much of the later tau propagation. Since higher neuronal activity promotes tau release and propagation [10, 11], the coupled SIS-QIF simulations presented herein suggest two possible tipping mechanisms. First, amyloid-induced hyperactivity could push tau pathology across a finite-amplitude invasion boundary, producing an abrupt transition from health to widespread pathological propagation. Second, if tau is already present at a low level, the same feedback could instead move the system from a low- to a high-tau state through endemic–endemic bistability. Tau pathology can indeed occur with little or no amyloid-β\beta in primary age-related tauopathy (PART) [53, 54]. The mechanism described here may explain how such tau pathology could become much more extensive after amyloid-β\beta accumulates. The nonmonotone, approximately quadratic effect of amyloid-β\beta on activity makes such state-dependent feedback particularly plausible. Amyloid-β\beta may therefore make the brain more amenable to subsequent tau propagation through its effect on neuronal activity, consistent with recent human imaging findings [55].

The spiking QIF–SIS system is bistable, and we show that external input to the neuronal dynamics can move it between low- and high-pathology states. The change persists after the input is removed. Neuronal stimulation has also altered protein pathology experimentally. In mouse models, 40-Hz optogenetic or visual stimulation reduced amyloid-β\beta levels [56], while auditory stimulation reduced amyloid and phosphorylated tau, and combined auditory and visual stimulation reduced amyloid more widely [57]. These interventions may also act through mechanisms which are not represented here. Nonetheless, our model suggests that suitably chosen neuronal stimulation could, in principle, prevent or halt pathological protein propagation by acting through activity–pathology feedback.

Limitations and outlook

Several limitations should be noted. The center-manifold calculation uses a node-level mean-field approximation and therefore neglects correlations between infection states. Such correlations are especially important on sparse or clustered networks and may shift both the invasion threshold and the saddle-node bifurcations. The calculation also assumes that disease-free activity approaches a stable fixed point. If activity is periodic or irregular, the growth of a small infection depends on the entire time-varying activity trajectory rather than on a single activity value. Moreover, the exclusion of periodic orbits and the global sign rule for multistability assume homogeneous dynamics on regular graphs. Their conclusions may change on irregular or directed networks, with sign-changing transmission multipliers, feedback coupling that depends on both source- and target-node states, or delays. Such extensions may support multistability with opposite-sign feedback, periodic orbits, and other dynamics outside the homogeneous reduction. Finally, because the disease-free state of the finite stochastic system is absorbing, the deterministic endemic branches represent long-lived quasistationary regimes whose switching and extinction times may depend on network size. The minimal examples and QIF models nevertheless reproduce the predicted multistable dynamics under the assumptions studied here.

Taken together, these results suggest that neuronal activity may be more than an accelerator of pathology propagation. Through its feedback with pathology, activity can reshape the dynamical landscape of disease, creating tipping points, hysteresis, and multiple stable pathology levels. Transient changes in activity may therefore produce lasting changes in disease progression. Applying this framework to longitudinal multimodal neuroimaging may reveal signatures of such tipping points in human disease and help determine whether targeted modulation of activity can redirect pathology toward a lower-spreading state.

Mean-field closure, center-manifold reduction, and numerical continuation

Markov generator and first-moment equation

Let X=(X1,…,XN)⊤∈{0,1}NX=(X_1,\ldots,X_N)^{\top}\in\{0,1\}^N. For Xi=0X_i=0, let Xi,+X^{i,+} denote the state obtained by replacing XiX_i by one. For Xi=1X_i=1, let Xi,−X^{i,-} denote the state obtained by replacing XiX_i by zero. To distinguish the random microscopic activity from its deterministic mean, write UU for the continuous component of the joint stochastic process and define u(t)=E[U(t)]u(t)=\mathbb{E}[U(t)]. Between infection and clearance jumps, UU obeys U˙=G(U,X)\dot{U}=G(U,X).

The generator of the joint piecewise-deterministic Markov process is

Lφ(X,U)=∑i=1N(1−Xi)λi(X,U)[φ(Xi,+,U)−φ(X,U)]+∑i=1NζXi[φ(Xi,−,U)−φ(X,U)]+∇Uφ(X,U)⊤G(U,X),(26)\mathcal{L}\varphi(X,U)=\sum_{i=1}^{N}(1-X_i)\lambda_i(X,U)\left[\varphi(X^{i,+},U)-\varphi(X,U)\right] +\sum_{i=1}^{N}\zeta X_i\left[\varphi(X^{i,-},U)-\varphi(X,U)\right] +\nabla_U\varphi(X,U)^{\top}G(U,X), \tag*{(26)}

where λi(X,U)=β∑j=1NWijΦ(Uj)Xj\lambda_i(X,U)=\beta\sum_{j=1}^{N}W_{ij}\Phi(U_j)X_j.

Choose φ(X,U)=Xi\varphi(X,U)=X_i. The continuous activity term vanishes because XiX_i does not depend on UU. The generator gives LXi=(1−Xi)λi(X,U)−ζXi\mathcal{L}X_i=(1-X_i)\lambda_i(X,U)-\zeta X_i. Dynkin’s formula therefore yields

ddtE[Xi]=β∑j=1NWijE[(1−Xi)Φ(Uj)Xj]−ζE[Xi].(27)\frac{d}{dt}\mathbb{E}[X_i]=\beta\sum_{j=1}^{N}W_{ij}\mathbb{E}\left[(1-X_i)\Phi(U_j)X_j\right]-\zeta\mathbb{E}[X_i]. \tag*{(27)}

The equation is not closed because it contains joint infection and activity moments. The node-level mean-field closure makes two distinct approximations. First, it factorizes the mixed moment as

E[(1−Xi)Φ(Uj)Xj]≈(1−xi)Φ(uj)xj,\mathbb{E}\left[(1-X_i)\Phi(U_j)X_j\right]\approx(1-x_i)\Phi(u_j)x_j,

where xi=E[Xi]x_i=\mathbb{E}[X_i] and uj=E[Uj]u_j=\mathbb{E}[U_j]. This closure neglects correlations between XiX_i, XjX_j, and UjU_j, and also makes the nonlinear activity approximation

E[Φ(Uj)]≈Φ(E[Uj])=Φ(uj),\mathbb{E}[\Phi(U_j)]\approx\Phi(\mathbb{E}[U_j])=\Phi(u_j),

which is exact for affine Φ\Phi but not in general. Substitution into (27) gives

x˙i=β(1−xi)∑j=1NWijΦ(uj)xj−ζxi.\dot{x}_i=\beta(1-x_i)\sum_{j=1}^{N}W_{ij}\Phi(u_j)x_j-\zeta x_i.

Writing this equation for all nodes gives

x˙=β(1−x)⊙Wdiag⁡(Φ(u))x−ζx.\dot{x}=\beta(1-x)\odot W\operatorname{diag}(\Phi(u))x-\zeta x.

Second, applying the generator to the activity coordinates gives the exact first-moment relation

ddtE[U]=E[G(U,X)].\frac{d}{dt}\mathbb{E}[U]=\mathbb{E}[G(U,X)].

The deterministic activity equation u˙=G(u,x)\dot{u} = \mathcal{G}(u,x) therefore assumes

E[G(U,X)]≈G(u,x).\mathbb{E}[\mathcal{G}(U,X)] \approx\mathcal{G}(u,x).

Because u=E[U]u = \mathbb{E}[U] and x=E[X]x = \mathbb{E}[X], this identity is exact when G\mathcal{G} is affine in (U,X)(U,X). For nonlinear G\mathcal{G}, it is a mean-field approximation justified when fluctuations and correlations are sufficiently weak. Together, these two assumptions close the deterministic system (3).

Center-manifold reduction at the epidemic threshold

Set u=a+yu = a + y, where G(a,0)=0\mathcal{G}(a,\mathbf{0}) = \mathbf{0}. The shifted deterministic system is

x˙=β(1−x)⊙Wdiag⁡(Φ(a+y))x−ζx,y˙=G(a+y,x).\begin{aligned} \dot{x} &= \beta(\mathbf{1}-x)\odot W\operatorname{diag}\bigl(\Phi(a+y)\bigr)x-\zeta x,\\ \dot{y} &= \mathcal{G}(a+y,x). \end{aligned}

Let

A=diag⁡(Φ(a)),JA=DuG(a,0),C=DxG(a,0).A = \operatorname{diag}(\Phi(a)), \qquad J_A = D_u\mathcal{G}(a,0), \qquad C = D_x\mathcal{G}(a,0).

The disease-free equilibrium is (x,y)=(0,0)(x,y) = (\mathbf{0},\mathbf{0}). Since JAJ_A is stable, the transcritical bifurcation occurs at

βc=ζρ0,ρ0=ρ(WA).\beta_c = \frac{\zeta}{\rho_0}, \qquad\rho_0 = \rho(WA).

Let r0r_0 and ℓ0\ell_0 be the right and left Perron eigenvectors of WAWA, normalized by ℓ0Tr0=1\ell_0^{\mathsf{T}}r_0 = 1, and define

Lc=βcWA−ζI,h=−JA−1Cr0.L_c = \beta_cWA-\zeta I, \qquad h=-J_A^{-1}Cr_0.

For the full state Z=(xT,yT)TZ=(x^{\mathsf{T}},y^{\mathsf{T}})^{\mathsf{T}}, the right and left nullvectors of the threshold Jacobian are

v=(r0h),q=(ℓ00),qTv=1.v=\begin{pmatrix}r_0\\h\end{pmatrix}, \qquad q=\begin{pmatrix}\ell_0\\0\end{pmatrix}, \qquad q^{\mathsf{T}}v=1.

Let μ=β−βc\mu=\beta-\beta_c, write the right-hand side as Z˙=F(Z;βc+μ)\dot{Z}=\mathcal{F}(Z;\beta_c+\mu), and choose the center coordinate

z=qTZ=ℓ0Tx.z=q^{\mathsf{T}}Z=\ell_0^{\mathsf{T}}x.

The center manifold theorem gives a locally invariant family Z=H(z,μ)=zv+O(z2,μz)Z=\mathcal{H}(z,\mu)=zv+O(z^2,\mu z). At the threshold we write its quadratic approximation as

H(z,0)=(zr0+z2ξzh+z2ψ)+O(z3),ℓ0Tξ=0.\mathcal{H}(z,0)=\begin{pmatrix}zr_0+z^2\xi\\zh+z^2\psi\end{pmatrix}+O(z^3), \qquad\ell_0^{\mathsf{T}}\xi=0.

Here ξ\xi and ψ\psi are the quadratic corrections to the infection and activity components, respectively. The reduced equation has the form

z˙=ρ0μz+c2z2+c3z3+O(μ2z,μz2,z4).\dot{z}=\rho_0\mu z+c_2z^2+c_3z^3+O(\mu^2z,\mu z^2,z^4).

The center-manifold invariance equation is

DzH(z,μ)z˙=F(H(z,μ);βc+μ).D_z\mathcal{H}(z,\mu)\dot{z}=\mathcal{F}\bigl(\mathcal{H}(z,\mu);\beta_c+\mu\bigr).

To determine the coefficients in the reduced equation, we substitute the center-manifold approximation into the invariance equation. For this purpose, define

D1[w]=diag⁡(Φ′(a)⊙w),D2[w1,w2]=diag⁡(Φ′′(a)⊙w1⊙w2),D_1[w]=\operatorname{diag}\bigl(\Phi'(a)\odot w\bigr), \qquad D_2[w_1,w_2]=\operatorname{diag}\bigl(\Phi''(a)\odot w_1\odot w_2\bigr),

and

QA(y,x)=12D2G(a,0)[(yx),(yx)].Q_A(y,x)=\frac{1}{2}D^2G(a,0)\left[\begin{pmatrix}y\\x\end{pmatrix},\begin{pmatrix}y\\x\end{pmatrix}\right].

Expanding the original system at β=βc\beta=\beta_c gives

x˙=Lcx−βcx⊙WAx+βcWD1[y]x−βcx⊙WD1[y]x+βc2WD2[y,y]x+O(∥(x,y)∥4),\dot{x}=L_cx-\beta_c x\odot WAx+\beta_cWD_1[y]x -\beta_cx\odot WD_1[y]x+\frac{\beta_c}{2}WD_2[y,y]x+O\left(\lVert(x,y)\rVert^4\right),
y˙=JAy+Cx+QA(y,x)+O(∥(x,y)∥3).\dot{y}=J_Ay+Cx+Q_A(y,x)+O\left(\lVert(x,y)\rVert^3\right).

We now match powers of zz and μ\mu. At order μz\mu z, projection with qTq^\mathsf{T} gives

ℓ0TWAr0=ρ0.\ell_0^\mathsf{T}WAr_0=\rho_0.

At quadratic order, let

n2=−ζr0⊙r0+βcWD1[h]r0n_2=-\zeta r_0\odot r_0+\beta_cWD_1[h]r_0

be the coefficient of z2z^2 in the infection equation along the critical direction. Projection with ℓ0T\ell_0^\mathsf{T} gives

c2=−ζℓ0T(r0⊙r0)+βcℓ0TWD1[h]r0.c_2=-\zeta\ell_0^\mathsf{T}(r_0\odot r_0)+\beta_c\ell_0^\mathsf{T}WD_1[h]r_0.

Let S={w∈RN:ℓ0Tw=0}S=\{w\in\mathbb{R}^N:\ell_0^\mathsf{T}w=0\}, and denote the inverse of the restriction of LcL_c to SS by Lc,S−1L_{c,S}^{-1}. This inverse exists because the zero eigenvalue of LcL_c is simple. Since ℓ0T(c2r0−n2)=0\ell_0^\mathsf{T}(c_2r_0-n_2)=0, the quadratic corrections are

ξ=Lc,S−1(c2r0−n2),ψ=JA−1[c2h−Cξ−QA(h,r0)].\begin{aligned} \xi&=L_{c,S}^{-1}(c_2r_0-n_2),\\ \psi&=J_A^{-1}[c_2h-C\xi-Q_A(h,r_0)]. \end{aligned}

It remains only to match the cubic terms. After substituting the known ξ\xi and ψ\psi, the cubic terms in the infection equation that contribute after projection are

n3=−βc[r0⊙WAξ+ξ⊙WAr0]+βcW[D1[h]ξ+D1[ψ]r0+12D2[h,h]r0]−βcr0⊙WD1[h]r0.\begin{aligned} n_3={}&-\beta_c[r_0\odot WA\xi+\xi\odot WAr_0]\\ &+\beta_cW\left[D_1[h]\xi+D_1[\psi]r_0+\frac{1}{2}D_2[h,h]r_0\right]\\ &-\beta_cr_0\odot WD_1[h]r_0. \end{aligned}

Projecting the matched cubic equation with ℓ0T\ell_0^\mathsf{T} and using ℓ0Tξ=0\ell_0^\mathsf{T}\xi=0 gives

c3=ℓ0Tn3.c_3=\ell_0^\mathsf{T}n_3.

which is (11) in the main text and completes the cubic center-manifold reduction. The coefficients c2c_2 and c3c_3 are used for the center-manifold predictions in Figure 1(a,b). Panel (c) also uses c4c_4. We obtain it numerically by first determining the cubic center-manifold correction from the unprojected z3z^3 terms and then matching the z4z^4 terms in the same invariance equation. We omit this standard but lengthy fourth-order calculation.

Numerical continuation and stability for Figure 1

For Figure 1, we use a fixed undirected Erdős–Rényi realization G(20,0.2)G(20, 0.2) and normalize its adjacency matrix so that ρ(W)=1\rho(W) = 1. We set ζ=1\zeta= 1 and use, for panels (a–c), respectively,

Φ1(u)=1+u2,Φ2(u)=(1+u)2,Φ3(u)=1+3u5+5u22.\Phi_1(u) = 1 + \frac{u}{2}, \qquad\Phi_2(u) = (1 + u)^2, \qquad\Phi_3(u) = 1 + \frac{3u}{5} + \frac{5u^2}{2}.

For each k=1,2,3k = 1, 2, 3, the full system is

x˙=β(1−x)⊙Wdiag⁡(Φk(u))x−x,u˙=−u+x.\dot{x} = \beta(1 - x) \odot W \operatorname{diag}(\Phi_k(u))x - x, \qquad\dot{u} = -u + x.

Thus every equilibrium satisfies u=xu = x, and βc=1\beta_c = 1.

We parameterize each endemic branch by z=ℓ0Tx>0z = \ell_0^\mathsf{T}x > 0. For each coupling function Φk\Phi_k and each prescribed zz, we solve the N+1N + 1 equations

(β(1−x)⊙Wdiag⁡(Φk(x))x−xℓ0Tx−z)=0\begin{pmatrix} \beta(1 - x) \odot W \operatorname{diag}(\Phi_k(x))x - x \\ \ell_0^\mathsf{T}x - z \end{pmatrix} = 0

for (x,β)(x, \beta). The continuation starts near the threshold with z0=2×10−5z_0 = 2 \times10^{-5}, using x=z0r0x = z_0 r_0 and β=βc\beta= \beta_c as the initial guess. For each kk, we use 700 equally spaced values of zz from z0z_0 to zmax⁡,k=0.55,0.78,0.62z_{\max,k} = 0.55, 0.78, 0.62, respectively. At each subsequent value, the preceding equilibrium, with its infection vector rescaled by the ratio of successive zz-values, initializes a Levenberg–Marquardt solve of the augmented equations above. The solver tolerances are 10−1210^{-12}. This continuation in zz passes through folds in β\beta; the disease-free branch is added separately.

The black dashed comparison curves are evaluated independently of the center-manifold reduced equation, using c2c_2, c3c_3 in panels (a,b) and c2c_2, c3c_3, c4c_4 in panel (c).

At every computed equilibrium, we evaluate the full 2N2N-dimensional Jacobian. A branch is classified as stable when all eigenvalues have negative real parts and unstable otherwise; these classifications give the solid and dashed curves in Figure 1(a–c). The trajectories in panels (d–f) are direct integrations of the full deterministic system from the lower and higher initial conditions specified in the figure caption.

Nonlinear feedback on regular graphs

Exclusion of periodic orbits

For the homogeneous system (14), write

f(x,y)=x[βρ(1−x)Φ(a+y)−ζ],g(x,y)=−y+H(x).f(x,y) = x\left[\beta\rho(1-x)\Phi(a+y)-\zeta\right], \qquad g(x,y) = -y + H(x).

On the half-plane x>0x > 0, choose the Dulac function B(x,y)=1/xB(x,y) = 1/x. Then

∂∂x(Bf)+∂∂y(Bg)=∂∂x[βρ(1−x)Φ(a+y)−ζ]+∂∂y[−y+H(x)x]=−βρΦ(a+y)−1x<0.\begin{aligned} \frac{\partial}{\partial x}(Bf) + \frac{\partial}{\partial y}(Bg) = \frac{\partial}{\partial x}\left[\beta\rho(1-x)\Phi(a+y)-\zeta\right] + \frac{\partial}{\partial y}\left[\frac{-y+H(x)}{x}\right] \\ &= -\beta\rho\Phi(a+y) - \frac{1}{x} < 0. \end{aligned}

The Bendixson–Dulac criterion therefore rules out a periodic orbit in any simply connected physical region with x>0x > 0 and Φ>0\Phi> 0. Here Φ>0\Phi> 0 means that activity may strengthen or weaken an existing transmission edge without reversing its sign, which is a basic assumption for the spreading processes considered here. If an application allows activity to drive Φ\Phi through zero and flip effective edge signs, thereby changing the signed network structure, the divergence need not remain negative and this argument no longer rules out periodic orbits.

Exact polynomial examples and numerical verification

The following are the polynomial transmission functions used in Figure 2. We set a=0a=0, ρ=ζ=1\rho=\zeta=1, and H(x)=xH(x)=x in (14). At an endemic equilibrium y=xy=x. Writing α:=ζ/(βρ)=1/β\alpha:= \zeta/(\beta\rho)=1/\beta, the equilibrium condition is (1−x)Φ(x)−α=0(1-x)\Phi(x)-\alpha=0. The factorizations below therefore display all endemic infection levels for the corresponding values of α\alpha.

For the quadratic example, take

Φ2(y)=1−157y+507y2,α2=2735.\Phi_2(y)=1-\frac{15}{7}y+\frac{50}{7}y^2,\qquad\alpha_2=\frac{27}{35}.

Then

(1−x)Φ2(x)−α2=−507(x−110)(x−25)(x−45).(28)(1-x)\Phi_2(x)-\alpha_2=-\frac{50}{7}\left(x-\frac{1}{10}\right)\left(x-\frac{2}{5}\right)\left(x-\frac{4}{5}\right). \tag*{(28)}

For the cubic example, take

Φ3(y)=1+12y−3409y2+4009y3,α3=4225.\Phi_3(y)=1+12y-\frac{340}{9}y^2+\frac{400}{9}y^3,\qquad\alpha_3=\frac{42}{25}.

The exact factorization is

(1−x)Φ3(x)−α3=−4009(x−110)(x−310)×(x−35)(x−1720).(29)\begin{aligned} (1-x)\Phi_3(x)-\alpha_3&=-\frac{400}{9}\left(x-\frac{1}{10}\right)\left(x-\frac{3}{10}\right)\\ &\quad\times\left(x-\frac{3}{5}\right)\left(x-\frac{17}{20}\right). \tag*{(29)} \end{aligned}

For the quartic example, take

Φ4(y)=1−140701079y+722001079y2−1320001079y3+800001079y4,α4=2431079.(30)\Phi_4(y)=1-\frac{14070}{1079}y+\frac{72200}{1079}y^2-\frac{132000}{1079}y^3+\frac{80000}{1079}y^4,\qquad\alpha_4=\frac{243}{1079}. \tag*{(30)}

Then

(1−x)Φ4(x)−α4=−800001079(x−110)(x−14)(x−1120)×(x−45)(x−1920).(31)\begin{aligned} (1-x)\Phi_4(x)-\alpha_4&=-\frac{80000}{1079}\left(x-\frac{1}{10}\right)\left(x-\frac{1}{4}\right)\left(x-\frac{11}{20}\right)\\ &\quad\times\left(x-\frac{4}{5}\right)\left(x-\frac{19}{20}\right). \tag*{(31)} \end{aligned}

All three transmission multipliers are positive for 0≤y≤10\le y\le1, so the examples satisfy the physical assumption Φ>0\Phi>0. For each polynomial, we numerically computed the equilibria and their stability and simulated (14) using an adaptive eighth-order Runge–Kutta method.

Monostability of the homogeneous QIF system

At fixed homogeneous infection xx, setting the two QIF population equations to zero gives

Q(r;x)=Δ24π2r2+ηˉ+H(x)+(Jloc+K)r−π2r2=0,v=−Δ2πr.(32)Q(r;x)=\frac{\Delta^2}{4\pi^2r^2}+\bar{\eta}+H(x)+(J_{\mathrm{loc}}+K)r-\pi^2r^2=0,\qquad v=-\frac{\Delta}{2\pi r}. \tag*{(32)}

Here we used ∑jWijr=r\sum_j W_{ij}r=r on the homogeneous regular graph. For Jloc+K≤0J_{\mathrm{loc}}+K\le0,

∂rQ=−Δ22π2r3+Jloc+K−2π2r<0(33)\partial_r Q=-\frac{\Delta^2}{2\pi^2r^3}+J_{\mathrm{loc}}+K-2\pi^2r<0 \tag*{(33)}

for every r>0r>0. Moreover, Q(r;x)Q(r;x) tends to +∞+\infty as r→0+r\to0^{+} and to −∞-\infty as r→∞r\to\infty. It therefore has exactly one positive root r∗(x)r_{*}(x) for every fixed xx. At the corresponding equilibrium, the neuronal Jacobian satisfies

tr⁡J∗=4v∗ε<0,det⁡J∗=−2r∗ε2∂rQ(r∗;x)>0,\operatorname{tr} J_{*}=\frac{4v_{*}}{\varepsilon}<0,\qquad\det J_{*}=-\frac{2r_{*}}{\varepsilon^{2}}\partial_{r}Q(r_{*};x)>0,

so the equilibrium is locally asymptotically stable. Moreover, the Dulac function B(r,v)=r−2B(r,v)=r^{-2} gives

∂r(Br˙)+∂v(Bv˙)=−2Δετr3<0,\partial_{r}(B\dot{r})+\partial_{v}(B\dot{v})=-\frac{2\Delta}{\varepsilon\tau r^{3}}<0,

excluding periodic orbits in the physical half-plane r>0r>0. Finally, differentiating the stationary equation gives r∗′(x)=−H′(x)/∂rQr_{*}'(x)=-H'(x)/\partial_{r}Q, which has the same sign as H′(x)H'(x) and hence, for H(x)=κSx(x)H(x)=\kappa S_{x}(x) with increasing SxS_{x}, the same sign as κ\kappa.

Simulation of the microscopic QIF–SIS system

Numerical method

Each node carries a circular theta-neuron phase θi\theta_i, an exponentially filtered spike variable sis_i, and a binary SIS state Xi∈{0,1}X_i\in\{0,1\}. Infection changes the neuronal current by κXi\kappa X_i, while the filtered firing of an infected source changes its outward transmission through 1+δSr(si)1+\delta S_r(s_i). We simulate these coupled processes using a split-step scheme of the type used for coupled neuronal ODE and Markov-jump models. Over each short interval, we first advance the neuronal variables with infection fixed and then sample the SIS jumps with the updated filtered firing fixed.

We use the fixed splitting step Δt=0.002\Delta t=0.002. Let XnX^n, θn\theta^n, and sns^n denote the states at the beginning of [tn,tn+Δt][t_n,t_n+\Delta t]. Holding XnX^n and sns^n fixed, we calculate the neuronal current

Iin=ηi+κXin+J∑jWijsjn+Iiext(tn).I_i^n=\eta_i+\kappa X_i^n+J\sum_j W_{ij}s_j^n+I_i^{\mathrm{ext}}(t_n).

Writing f(θ,I)=1−cos⁡θ+(1+cos⁡θ)If(\theta,I)=1-\cos\theta+(1+\cos\theta)I, we advance the phase by the explicit midpoint rule

θin+1=θin+Δtεf(θin+Δt2εf(θin,Iin),Iin).\theta_i^{n+1}=\theta_i^n+\frac{\Delta t}{\varepsilon}f\left(\theta_i^n+\frac{\Delta t}{2\varepsilon}f(\theta_i^n,I_i^n),I_i^n\right).

We set Nin=1N_i^n=1 if this midpoint step carries the phase of neuron ii to or beyond the spike value θi=π\theta_i=\pi, and Nin=0N_i^n=0 otherwise. The filtered firing is then updated as

sin+1=exp⁡(−Δtτs)sin+ετsNin.s_i^{n+1}=\exp\left(-\frac{\Delta t}{\tau_s}\right)s_i^n+\frac{\varepsilon}{\tau_s}N_i^n.

We then keep sn+1s^{n+1} fixed and simulate the SIS process over the same interval using the direct Gillespie method [34]. For a susceptible node, the infection rate is

λi(X;sn+1)=βτX∑jWij(1+δSr(sjn+1))Xj.\lambda_i(X;s^{n+1})=\frac{\beta}{\tau_X}\sum_j W_{ij}\left(1+\delta S_r(s_j^{n+1})\right)X_j.

Thus, for the current binary state XX, the total event rate at node ii is

ai(X;sn+1)=(1−Xi)λi(X;sn+1)+XiζτX,a_i(X;s^{n+1})=(1-X_i)\lambda_i(X;s^{n+1})+X_i\frac{\zeta}{\tau_X},

where the first term is the infection rate when node ii is susceptible and the second is the recovery rate when it is infected. Let a0=∑iaia_0 = \sum_i a_i. We draw a waiting time from an exponential distribution with rate a0a_0. If it exceeds the time remaining in the interval, no further event occurs. Otherwise, node ii is selected with probability ai/a0a_i/a_0 and its state is switched. We then recompute the SIS event rates for the updated binary state and repeat only this Gillespie event calculation until the interval is exhausted; θn+1\theta^{n+1} and sn+1s^{n+1} remain fixed throughout this SIS substep. The resulting state is Xn+1X^{n+1}. The next splitting interval starts from (Xn+1,θn+1,sn+1)(X^{n+1},\theta^{n+1},s^{n+1}), and the same neuronal and SIS substeps are repeated.

The neuronal excitabilities are the deterministic Lorentzian quantiles

ηi=ηˉ+Δtan⁡[π(iN+1−12)],i=1,…,N.\eta_i = \bar{\eta} + \Delta\tan\left[\pi\left(\frac{i}{N+1}-\frac{1}{2}\right)\right], \qquad i=1,\ldots,N.

These are values of the inverse Lorentzian cumulative distribution function at the evenly spaced probabilities i/(N+1)i/(N+1). Choosing them deterministically rather than drawing them randomly makes the finite network represent the intended Lorentzian distribution evenly and eliminates sample-to-sample variation in the excitability values. We randomly permute this fixed set across the nodes to avoid correlations with network structure. Initially, each θi\theta_i is drawn uniformly from [−π,π][-\pi,\pi] and si=0.01s_i = 0.01. The initially infected nodes are selected uniformly at random, with their number specified separately for each experiment.

The spatial graph is generated once, with node positions sampled independently from a uniform distribution over the unit square and edges between nodes separated by at most 0.155. For Figure 5 and Figure 6, the two unforced trajectories are integrated to t=1200t = 1200 and the input-driven trajectory to t=5500t = 5500, with all trajectories sampled every 0.5.

Control with opposite-sign coupling

Figure 6 showed that, at β=0.76\beta= 0.76 and with same-sign feedback (κ,δ)=(3,12)(\kappa,\delta) = (3,12), a transient positive neuronal input can drive a 50-node infection seed into a high-endemic state which persists after the input is removed. The following control tests whether this persistence is merely slow stochastic decay. Recall that κ\kappa determines how infection changes the neuronal current, whereas δ\delta determines how neuronal firing changes transmission. In Figure 7, we repeat the experiment with the same graph and stochastic initialization but reverse κ\kappa from 33 to −3-3, while keeping δ=12\delta= 12. Infection therefore suppresses firing, whereas firing still promotes transmission, so the two coupling directions have opposite signs. We apply the stronger input Iiext=5I_i^{\mathrm{ext}} = 5 for 0≤t<5000 \le t < 500, which raises the prevalence to 0.687, comparable with the high-endemic range in the same-sign example. Once the input is removed, firing returns to its baseline level and infection decays to absorption at t=2167t = 2167.

References

  1. [1]Thilo Gross, Carlos J. Dommar D’Lima, and Bernd Blasius. Epidemic Dynamics on an Adaptive Network. Physical Review Letters, 96(20):208701, May 2006.
  2. [2]Kathinka Frieswijk, Lorenzo Zino, Mengbin Ye, Alessandro Rizzo, and Ming Cao. A Mean-Field Analysis of a Network Behavioral–Epidemic Model. IEEE Control Systems Letters, 6:2533–2538, 2022.
  3. [3]Massimo A. Achterberg and Mattia Sensi. A minimal model for adaptive SIS epidemics. Nonlinear Dynamics, 111(13):12657–12670, July 2023.
  4. [4]Tyrus Berry, Matthew Ferrari, Timothy Sauer, Steven J. Greybrush, Donald Ebeigbe, Andrew J. Whalen, and Steven J. Schiff. Existence and stability of equilibria in infectious disease dynamics with behavioral feedback. Physical Review. E, 111(1-1):014317, January 2025.
  5. [5]Tianyu Li, Yipeng Hu, Xuening Li, Xueqin Wang, Lijian Yang, and Ya Jia. Activity feedback and hyperedge overlap shape bistability in higher-order networks. Physical Review E, 114(2):024303, August 2026.
  6. [6]L. Böttcher, O. Woolley-Meza, N. a. M. Araújo, H. J. Herrmann, and D. Helbing. Disease-induced resource constraints can trigger explosive epidemics. Scientific Reports, 5(1):16571, November 2015.
  7. [7]Mehrdad Farajtabar, Yichen Wang, Manuel Gomez Rodriguez, Shuang Li, Hongyuan Zha, and Le Song. COEVOLVE: A Joint Point Process Model for Information Diffusion and Network Co-evolution. In Advances in Neural Information Processing Systems, volume 28. Curran Associates, Inc., 2015.
  8. [8]Mathias Jucker and Lary C. Walker. Self-propagation of pathogenic protein aggregates in neurodegenerative diseases. Nature, 501(7465):45–51, September 2013.
  9. [9]Michel Goedert. Alzheimer’s and Parkinson’s diseases: The prion concept in relation to assembled Aβ, tau, and α-synuclein. Science, 349(6248):1255555, August 2015.
  10. [10]Amy M Pooler, Emma C Phillips, Dawn H W Lau, Wendy Noble, and Diane P Hanger. Physiological release of endogenous tau is stimulated by neuronal activity. EMBO reports, 14(4):389–394, April 2013.DOI
  11. [11]Jessica W. Wu, S. Abid Hussaini, Isle M. Bastille, Gustavo A. Rodriguez, Ana Mrejeru, Kelly Rilett, David W. Sanders, Casey Cook, Hongjun Fu, Rick A. C. M. Boonen, Mathieu Herman, Eden Nahmani, Sheina Emrani, Y. Helen Figueroa, Marc I. Dimond, Catherine L. Clelland, Selina Wray, and Karen E. Duff. Neuronal activity enhances tau propagation and tau pathology in vivo. Nature Neuroscience, 19(8):1085–1092, August 2016.
  12. [12]Marc Aurel Busche, Gerhard Eichhoff, Helmuth Adelsberger, Dorothee Abramowski, Karl-Heinz Wiederhold, Christian Haass, Matthias Staufenbiel, Arthur Konnert, and Olga Garaschuk. Clusters of Hyperactive Neurons Near Amyloid Plaques in a Mouse Model of Alzheimer’s Disease. Science, 321(5896):1686–1689, September 2008.
  13. [13]Ashish Raj, Amy Kuceyeski, and Michael Weiner. A network diffusion model of disease progression in dementia. Neuron, 73(6):1204–1215, March 2012.
  14. [14]Jacob W. Vogel, Yasser Iturria-Medina, Olof T. Strandberg, Ruben Smith, Elizabeth Levitis, Alan C. Evans, and Oskar Hansson. Spread of pathological tau proteins through communicating neurons in human Alzheimer’s disease. Nature Communications, 11(1):1–15, May 2020.
  15. [15]Christoffer G. Alexandersen, Willem de Haan, Christian Bick, and Alain Goriely. A multi-scale model explains oscillatory slowing and neuronal hyperactivity in Alzheimer’s disease. Journal of The Royal Society Interface, 20(198):20220607, January 2023.
  16. [16]Christoffer G. Alexandersen, Alain Goriely, and Christian Bick. Neuronal activity induces symmetry breaking in neurodegenerative disease spreading. Journal of Mathematical Biology, 89(1):3, May 2024.
  17. [17]Jesús Cabrera-Álvarez, Leon Stefanovski, Leon Martin, Gianluca Susi, Fernando Maestú, and Petra Ritter. A Multiscale Closed-Loop Neurotoxicity Model of Alzheimer’s Disease Progression Explains Functional Connectivity Alterations. eNeuro, 11(4):ENEURO.0345–23.2023, April 2024.
  18. [18]Christoffer G. Alexandersen, Suman S. Kulkarni, Jessica T. Davis, Sebastian N. Roemer-Cassiano, Nicolai Franzmeier, and Dani S. Bassett. Activity-dependent epidemic spreading in multimultiscale brain networks predicts Alzheimer’s disease progression, August 2026. arXiv:2608.12647 [q-bio.NC] version: 1.arxiv.org/abs/2608.12647
  19. [19]Ana Lajmanovich and James A. Yorke. A deterministic model for gonorrhea in a nonhomogeneous population. Mathematical Biosciences, 28(3):221–236, January 1976.
  20. [20]Piet Van Mieghem, Jasmina Omic, and Robert Kooij. Virus Spread in Networks. IEEE/ACM Transactions on Networking, 17(1):1–14, February 2009.
  21. [21]István Z. Kiss, Joel C. Miller, and Péter L. Simon. Mathematics of Epidemics on Networks: From Exact to Approximate Models. Springer, June 2017.
  22. [22]P. van den Driessche and James Watmough. A simple SIS epidemic model with a backward bifurcation. Journal of Mathematical Biology, 40(6):525–540, June 2000.
  23. [23]Damián H. Zanette and Sebastián Risau-Gusmán. Infection Spreading in a Population with Evolving Contacts. Journal of Biological Physics, 34(1-2):135–148, April 2008.
  24. [24]M. R. Razvan and S. Yasaman. Global analysis of an SAIS model. Journal of Biological Dynamics, 6(2):457–474, March 2012.
  25. [25]Iacopo Iacopini, Giovanni Petri, Alain Barrat, and Vito Latora. Simplicial models of social contagion. Nature Communications, 10(1):2485, June 2019.
  26. [26]Yi-Jie Huang and Chun-Hsien Li. Backward bifurcation and stability analysis of a network-based SIS epidemic model with saturated treatment function. Physica A: Statistical Mechanics and its Applications, 527:121407, August 2019.
  27. [27]Xiaoyan Wang and Junyuan Yang. A Bistable Phenomena Induced by a Mean-Field SIS Epidemic Model on Complex Networks: A Geometric Approach. Frontiers in Physics, 9, June 2021.
  28. [28]S. Gómez, A. Arenas, J. Borge-Holthoefer, S. Meloni, and Y. Moreno. Discrete-time Markov chain approach to contact-based disease spreading in complex networks. Europhysics Letters, 89(3):38009, February 2010.
  29. [29]Sergio Gómez, Jesús Gómez-Gardeñes, Yamir Moreno, and Alex Arenas. Nonperturbative heterogeneous mean-field approach to epidemic spreading in complex networks. Physical Review E, 84(3):036105, September 2011.
  30. [30]Deepayan Chakrabarti, Yang Wang, Chenxi Wang, Jurij Leskovec, and Christos Faloutsos. Epidemic thresholds in real networks. ACM Transactions on Information and System Security (TISSEC), 10(4):1:1–1:26, January 2008.
  31. [31]Ernest Montbrió, Diego Pazó, and Alex Roxin. Macroscopic Description for Networks of Spiking Neurons. Physical Review X, 5(2):021028, June 2015.
  32. [32]Daniele Avitabile, Mathieu Desroches, and G. Bard Ermentrout. Cross-scale excitability in networks of quadratic integrate-and-fire neurons. PLOS Computational Biology, 18(10):e1010569, October 2022.
  33. [33]Pavol Bauer, Stefan Engblom, Sanja Mikulovic, and Aleksandar Senek. Multiscale modelling via split-step methods in neural firing. Mathematical and Computer Modelling of Dynamical Systems, 24(4):426–445, July 2018.
  34. [34]Daniel T. Gillespie. Exact stochastic simulation of coupled chemical reactions. The Journal of Physical Chemistry, 81(25):2340–2361, December 1977.
  35. [35]Christoffer G. Alexandersen, Georgia S. Brennan, Julia K. Brynildsen, Michael X. Henderson, Yasser Iturria-Medina, and Dani S. Bassett. Network Models of Neurodegeneration: Bridging Neuronal Dynamics and Disease Progression. IEEE Reviews in Biomedical Engineering, 19:140–158, 2026.
  36. [36]N. Perra, B. Goncalves, R. Pastor-Satorras, and A. Vespignani. Activity driven modeling of time varying networks. Scientific Reports, 2(1):469, June 2012.
  37. [37]Michele Starnini and Romualdo Pastor-Satorras. Temporal percolation in activity-driven networks. Physical Review E, 89(3):032807, March 2014.
  38. [38]Lorenzo Zino, Alessandro Rizzo, and Maurizio Porfiri. Continuous-Time Discrete-Distribution Theory for Activity-Driven Networks. Physical Review Letters, 117(22):228302, November 2016.
  39. [39]Jacopo Pozzana, Kaiyuan Sun, and Nicola Perra. Epidemic spreading on activity-driven networks with attractiveness. Physical Review. E, 96(4):042310, October 2017.
  40. [40]Suyu Liu, Nicola Perra, Márton Karsai, and Alessandro Vespignani. Controlling Contagion Processes in Activity Driven Networks. Physical Review Letters, 112(11):118702, March 2014.
  41. [41]Alessandro Rizzo, Mattia Frasca, and Maurizio Porfiri. Effect of individual behavior on epidemic spreading in activity-driven networks. Physical Review E, 90(4):042801, October 2014.
  42. [42]Masaki Ogura, Victor M. Preciado, and Naoki Masuda. Optimal Containment of Epidemics over Temporal Activity-Driven Networks. SIAM Journal on Applied Mathematics, 79(3):986–1006, January 2019.
  43. [43]Marco Mancastroppa, Raffaella Burioni, Vittoria Colizza, and Alessandro Vezzani. Active and inactive quarantine in epidemic spreading on adaptive activity-driven networks. Physical Review E, 102(2):020301, August 2020.
  44. [44]Qihui Wu, Muhammad A. Shaikh, Emily S. Meymand, Bin Zhang, Kelvin C. Luk, John Q. Trojanowski, and Virginia M.-Y. Lee. Neuronal activity modulates alpha-synuclein aggregation and spreading in organotypic brain slice cultures and in vivo. Acta Neuropathologica, 140(6):831–849, December 2020.
  45. [45]Marc Aurel Busche, Xiaowei Chen, Horst A. Henning, Julia Reichwald, Matthias Staufenbiel, Bert Sakmann, and Arthur Konnerth. Critical role of soluble amyloid-β for early hippocampal hyperactivity in a mouse model of Alzheimer’s disease. Proceedings of the National Academy of Sciences, 109(22):8740–8745, May 2012.
  46. [46]Marc Aurel Busche, Susanne Wegmann, Simon Dujardin, Caitlin Commins, Julia Schiantarelli, Naomi Klickstein, Tarun V. Kamath, George A. Carlson, Israel Nelken, and Bradley T. Hyman. Tau impairs neural circuits, dominating amyloid-β effects, in Alzheimer models in vivo. Nature Neuroscience, 22(1):57–64, January 2019.
  47. [47]Samuel S. Harris, Fred Wolf, Bart De Strooper, and Marc Aurel Busche. Tipping the Scales: Peptide-Dependent Dysregulation of Neural Circuit Dynamics in Alzheimer’s Disease. Neuron, 107(3):417–435, August 2020.
  48. [48]Daniela Puzzo, Lucia Privitera, Elena Leznik, Mauro Fà, Agnieszka Staniszewski, Agostino Palmeri, and Ottavio Arancio. Picomolar amyloid-beta positively modulates synaptic plasticity and memory in hippocampus. The Journal of Neuroscience: The Official Journal of the Society for Neuroscience, 28(53):14537–14545, December 2008.
  49. [49]Jorge J. Palop and Lennart Mucke. Amyloid-β–induced neuronal dysfunction in Alzheimer’s disease: from synapses toward neural networks. Nature Neuroscience, 13(7):812–818, July 2010.
  50. [50]Bernard J. Hanseeuw, Rebecca A. Betensky, Heidi L. L. Jacobs, Aaron P. Schultz, Jorge Sepulcre, J. Alex Becker, Danielle M. Orozco Cosio, Michelle Farrell, Yakeel T. Quiroz, Elizabeth C. Mormino, Rachel F. Buckley, Kathryn V. Papp, Rebecca A. Amariglio, Ilse Dewachter, Adrian Ivanoiu, Willem Huijbers, Trey Hedden, Gad A. Marshall, Jasmeer P. Chhatwal, Dorene M. Rentz, Reisa A. Sperling, and Keith Johnson. Association of Amyloid and Tau With Cognition in Preclinical Alzheimer Disease: A Longitudinal Study. JAMA Neurology, 76(8):915–924, August 2019.
  51. [51]Niklas Mattsson-Carlgren, Emilie Andersson, Shorena Janelidze, Rik Ossenkoppele, Philip Insel, Olof Strandberg, Henrik Zetterberg, Howard J. Rosen, Gil Rabinovici, Xiyun Chai, Kaj Blennow, Jeffrey L. Dage, Erik Stomrud, Ruben Smith, Sebastian Palmqvist, and Oskar Hansson. Aβ deposition is associated with increases in soluble and phosphorylated tau that precede a positive Tau PET in Alzheimer’s disease. Science Advances, 6(16):eaaz2387, April 2020.
  52. [52]Justin S. Sanchez, Bernard J. Hanseeuw, Francisco Lopera, Reisa A. Sperling, Ana Baena, Yamile Bocanegra, David Aguillon, Edmarie Guzmán-Vélez, Emmanuelle Pardilla-Delgado, Liliana Ramirez-Gomez, Clara Vila-Castelar, Jairo E. Martinez, Joshua T. Fox-Fuller, Claudia Ramos, Martin Ochoa-Escudero, Sergio Alvarez, Heidi I. L. Jacobs, Aaron P. Schultz, Jennifer R. Gatchel, J. Alex Becker, Samantha R. Katz, Danielle V. Mayblyum, Julie C. Price, Eric M. Reiman, Keith A. Johnson, and Yakeel T. Quiroz. Longitudinal amyloid and tau accumulation in autosomal dominant Alzheimer’s disease: findings from the Colombia-Boston (COLBOS) biomarker study. Alzheimer’s Research & Therapy, 13(1):27, January 2021.
  53. [53]John F. Crary, John Q. Trojanowski, Julie A. Schneider, Jose F. Abisambra, Erin L. Abner, Irina Alafuzoff, Steven E. Arnold, Johannes Attems, Thomas G. Beach, Eileen H. Bigio, Nigel J. Cairns, Dennis W. Dickson, Marla Gearing, Lea T. Grinberg, Patrick R. Hof, Bradley T. Hyman, Kurt Jellinger, Gregory A. Jicha, Gabor G. Kovacs, David S. Knopman, Julia Kofler, Walter A. Kukull, Ian R. Mackenzie, Eliezer Masliah, Ann McKee, Thomas J. Montine, Melissa E. Murray, Janna H. Neltner, Ismael Santa-Maria, William W. Seeley, Alberto Serrano-Pozo, Michael L. Shelanski, Thor Stein, Masaki Takao, Dietmar R. Thal, Jonathan B. Toledo, Juan C. Troncoso, Jean Paul Vonsattel, Charles L. White, Thomas Wisniewski, Randall L. Woltjer, Masahito Yamada, and Peter T. Nelson. Primary age-related tauopathy (PART): a common pathology associated with human aging. Acta Neuropathologica, 128(6):755–766, December 2014.
  54. [54]Keith A. Josephs, Melissa E. Murray, Nirubol Tosakulwong, Jennifer L. Whitwell, David S. Knopman, Mary M. Machulda, Stephen D. Weigand, Bradley F. Boeve, Kejal Kantarci, Leonard Petrucelli, Val J. Lowe, Clifford R. Jack, Ronald C. Petersen, Joseph E. Parisi, and Dennis W. Dickson. Tau aggregation influences cognition and hippocampal atrophy in the absence of beta-amyloid: a clinico-imaging-pathological study of primary age-related tauopathy (PART). Acta Neuropathologica, 133(5):705–715, May 2017.
  55. [55]Christoffer G. Alexandersen, Dani S. Bassett, Alain Goriely, and PavanJit Chaggar. Neuronal activity and amyloid-β promote tau seeding in the entorhinal cortex in Alzheimer’s disease. Brain: A Journal of Neurology, 149(6):1915–1928, June 2026.
  56. [56]Hunter F. Iaccarino, Annabelle C. Singer, Anthony J. Martorell, Andrii Rudenko, Fan Gao, Tyler Z. Gillingham, Hansruedi Mathys, Jinsoo Seo, Oleg Kritskiy, Fatema Abdurrob, Chinnakkaruppan Adaikkan, Rebecca G. Canter, Richard Rueda, Emery N. Brown, Edward S. Boyden, and Li-Huei Tsai. Gamma frequency entrainment attenuates amyloid load and modifies microglia. Nature, 540(7632):230–235, December 2016.
  57. [57]Anthony J. Martorell, Abigail L. Paulson, Ho-Jun Suk, Fatema Abdurrob, Gabrielle T. Drummond, Webster Guan, Jennie Z. Young, David Nam-Woo Kim, Oleg Kritskiy, Scarlett J. Barker, Vamsi Mangena, Stephanie M. Prince, Emery N. Brown, Kwanghun Chung, Edward S. Boyden, Annabelle C. Singer, and Li-Huei Tsai. Multi-sensory Gamma Stimulation Ameliorates Alzheimer’s-Associated Pathology and Improves Cognition. Cell, 177(2):256–271.e22, April 2019.

Paper details

vlnc:W_dj29dk5k
Version
v1 · latest
Published
Added
License
CC BY 4.0
  • Agricultural and Biological Sciences

Not recorded for this paper

View statement
Looking for related papers…
Contents