Biophysical Tau Propagation Models

• Biophysical tau propagation models are mathematical frameworks that simulate tau misfolding, local reaction kinetics, and network-mediated diffusion to capture Alzheimer’s pathology.
• They combine reaction–diffusion systems with graph-based transport to model anisotropic diffusion, active transport, and nonlocal effects, calibrated with imaging data.
• These models inform personalized forecasting and therapeutic strategies by quantitatively linking microscopic tau dynamics with clinical biomarkers.

Aggregates of misfolded tau protein drive the progression of Alzheimer's disease (AD) through autocatalytic propagation in neural tissue and via long-range network transport. Biophysical tau propagation models mathematically formalize these processes, aiming to quantitatively reconcile microscopic cellular mechanisms, mesoscale connectivity, and macroscopic disease spread observed in neuroimaging and pathology. These models encompass reaction–diffusion systems with varying complexity, incorporating local misfolding kinetics, anisotropic (often network-mediated) transport, active cellular mechanisms, coupling to amyloid and atrophy, nonlocal and stochastic effects, and, in advanced frameworks, parametric inference directly from imaging data. The field has matured to support personalized inference of spatiotemporal propagation patterns, hypothesis testing on biophysical driving factors, and integration with state-of-the-art computational and neural operator surrogates.

1. Mathematical Structures: Local Reaction–Diffusion and Network Models

Biophysical tau propagation models are derived from two principal mathematical archetypes: local reaction–diffusion systems and network-based dynamical models.

A. Local Reaction–Diffusion PDEs

Tau concentration $c(\mathbf{x},t)$ in continuous brain tissue evolves according to spatially heterogeneous reaction–diffusion, often with anisotropy aligned to white-matter fibers: $$\frac{\partial c}{\partial t} = \nabla\cdot\left(\mathbf{D}(\mathbf{x})\,\nabla c\right) + f(c) \quad \mathrm{on}~\Omega\times(0,T]$$ where $\mathbf{D}(\mathbf{x})$ encodes axonal and extracellular diffusivity and $f(c)$ represents local conversion/clearance kinetics, frequently of Fisher–Kolmogorov or logistic form ($f(c) = \alpha c(1-c)$) (Zhang et al., 2023, Antonietti et al., 2024, 2012.19661, Kevrekidis et al., 2020).

B. Two-Species (Heterodimer) and Coupled Protein Models

An explicit representation of healthy ($p$) and misfolded ($q$) tau enables modeling of conversion and clearance: $$\begin{cases} \frac{\partial p}{\partial t} = \nabla\cdot\left(\mathbf{D}\nabla p\right) - k_1 p - k_{12} p q + k_0 \ \frac{\partial q}{\partial t} = \nabla\cdot\left(\mathbf{D}\nabla q\right) - \tilde{k}_1 q + k_{12} p q \end{cases}$$ Here, $k_0$ is the production of healthy tau, $k_1$, $\tilde{k}_1$ clearance, $k_{12}$ misfolding rate (Corti, 2024, Antonietti et al., 2024, Pal et al., 2021).

C. Network and Graph-Based Models

On the connectome, tau propagation is formulated as a set of ODEs/PDEs over brain regions indexed by $i$, with graph-Laplacian (diffusion on $W$ weighted by DTI) or directional edges: $$\frac{d x_{T,i}}{dt} = k_{pT} - k_{dT} x_{T,i} + F_{\mathrm{rxn}}(x_{A,i}) + d_T \sum_{j} L_{ij} x_{T,j}$$ $L = D - W$ is the graph Laplacian; $F_{\mathrm{rxn}}$ includes amyloid-induced seeding (Zhang et al., 2020, Zhang et al., 2022). The network transport model (NTM) further captures directed, active axonal transport and edge-specific mechanisms (Tora et al., 2024, Barron et al., 9 Mar 2026).

D. Nonlocal and Fractional Models

Nonlocal operators and fractional derivatives ($D^\alpha_t$) introduce spatial and temporal memory, capturing anomalous transport, astrocytic effects, or connectome-wide “seeding” (Pal et al., 2021, Pal et al., 2022).

2. Biophysical Mechanisms and Model Components

A. Reaction Kinetics

• Autocatalytic misfolding: Modeled as $f(c)$ (often cubic/empirically discovered) or $k_{12} p q$ (heterodimer models).
  • Discovered by PINN-informed symbolic regression (e.g., $f^-(c) = -c^3 + 0.62c^2 + 0.39c$ for healthy; $f^+(c) = 0.23c^3 - 1.34c^2 + 1.11c$ for AD) (Zhang et al., 2023).
• Amyloid–tau coupling: Modeled via terms such as $b_3 \tilde{u} v \tilde{v}$ or Hill-type activation, enabling primary/secondary tauopathy regimes (Kevrekidis et al., 2020, Pal et al., 2021, Zhang et al., 2022).
• Tissue atrophy coupling: Logistic mass-loss law for atrophy variable $g$ linked to tau concentration $c$ (Pederzoli et al., 2024).

B. Transport Mechanisms

• Anisotropic diffusion: $\mathbf{D}$ is constructed as $$d_{ext}\mathbf{I} + d_{axn}\mathbf{n}\otimes \mathbf{n}$$ with $d_{axn} \gg d_{ext}$ in white matter (2012.19661, Antonietti et al., 2024, Corti, 2024)
• Active transport (NTM): Explicit motor-driven advection along directed edges, including molecular feedback from tau–motor interactions (Tora et al., 2024, Barron et al., 9 Mar 2026).
• Nonlocal coupling: Connectome-based convolution or fractional Laplacian to model long-range/trans-synaptic effects (Pal et al., 2021, Pal et al., 2022).

C. Multi-layer and Modulatory Effects

• SC/FC multi-layer models: Parallel, bidirectionally coupled diffusion on anatomical (structural connectivity, SC) and functional (FC) graphs, with time- and region-varying dominance modulated by amyloid burden, genetics (APOE, CHUK/TMEM106B/MCL1/NOTCH1/TH), and disease stage (Dan et al., 23 Oct 2025).
• Astrocyte dual-role and memory: Feedback terms switch between clearance and inflammatory enhancement, while fractional time-derivatives slow propagation (“disease memory” effect) (Pal et al., 2022).

3. Calibration, Inference, and Numerical Schemes

A. Parameter Estimation

Parameters are obtained via imaging–data–driven calibration: joint optimization against longitudinal PET/MRI/DTI using global schemes (genetic algorithms, adjoint optimization) (Zhang et al., 2020, Zhang et al., 2022, Zhang et al., 2023). Key rates inferred include diffusion coefficients $(d_{axn}, d_T, \kappa)$, reaction/growth $(\alpha, k_{12}, k_{pT})$, and coupling strengths $(b_3, K_{AT})$.

B. Direct Model Discovery

PINNs and symbolic regression identify $f(c)$, enabling the data-driven selection of reaction kinetics absent prior biochemical specification (Zhang et al., 2023).

C. Numerical Methods

State-of-the-art schemes address multiscale geometry and numerical stiffness:

• Discontinuous Galerkin (DG)/PolyDG on polygonal/polytopal meshes (Antonietti et al., 2024, Pederzoli et al., 2024)
• IMEX Runge-Kutta and Crank–Nicolson for (semi-)implicit time integration (Antonietti et al., 2024, Corti, 2024)
• Sparse initial-condition recovery for network-ODEs via $\ell_0$-constrained inverse optimization (Wen et al., 2024)

D. Surrogate Modeling

Operator learning (Tau-BNO) reproduces full NTM trajectories with high fidelity ($R^2\approx 0.98$), reducing computational cost by orders of magnitude and enabling high-throughput inference (Barron et al., 9 Mar 2026).

4. Predictive Outcomes, Phenomenology, and Model Comparison

A. Traveling Waves and Spreading Patterns

Models predict sharp, anisotropic propagation fronts aligned with axonal connectivity, reproducing Braak-stage staging and region-specific delays. Fisher–Kolmogorov models produce sharp invasion fronts; heterodimer models can capture overshoot and damped foci if reaction eigenvalues are complex (Kevrekidis et al., 2020, Corti, 2024, Antonietti et al., 2024).

B. Sensitivity and Dominant Drivers

Global Sobol' analysis shows that spread speed and spatial extent are most sensitive to the net autocatalytic rate $\alpha=p_\Delta k_{12}$ and to anisotropy parameters ($d_{axn} \gg d_{ext}$). Clearance and baseline healthy-tau variances are subdominant but non-negligible (Corti, 2024). In multi-layer models, the dominance of SC or FC pathways shifts with disease progression and genetics (Dan et al., 23 Oct 2025).

C. Analytical Results

Constant-coefficient models yield explicit traveling-wave speeds, $c = 2\sqrt{D_\parallel \alpha}$, and stability domains can be assessed via linearization and Lyapunov criteria (Kevrekidis et al., 2020, Zhang et al., 2020, Zhang et al., 2022).

D. Mechanistic Insights

• NTM and active-transport models show that directionality biases staging patterns (anterograde vs. retrograde spread) and staging order; aggregation sequesters soluble pools, slowing propagation and shifting regional maxima (Tora et al., 2024, Barron et al., 9 Mar 2026).
• Astrocyte–tau models reveal that memory and inflammation impose spatial and temporal heterogeneity on vulnerability and recovery, with strong non-Markovian retardation of late-stage pathology (Pal et al., 2022).
• Nonlocal heterodimer and network models capture observed region-to-region heterogeneity, including the distinction between primary, secondary, and mixed tauopathy, and explain differential patterns of atrophy and cognitive decline (Pal et al., 2021, Dan et al., 23 Oct 2025).

5. Model Limitations, Assumptions, and Potential Extensions

A. Validation and Uncertainty

Biophysical rates are typically estimated via limited longitudinal imaging (often 2–3 time points), with fixed topology and groupwise homogeneous rates. This constrains model identifiability and precludes direct assessment of individual substrate variation, clearance, or PET off-target binding (Zhang et al., 2023, Wen et al., 2024).

B. Biological Simplifications

Most models assume time-invariant connectivity, neglecting progressive tissue damage or immune response feedback on physical network properties. Tau clearance (CSF efflux, proteolysis) is typically encoded as a simple first-order term. Active axonal transport and region-specific microenvironmental differences are incompletely represented (Tora et al., 2024, Pederzoli et al., 2024).

C. Further Directions

Proposed advances include subject- and region-specific heterogeneity in reaction polynomials and diffusivities, explicit coupling to amyloid, atrophy, and genetic risk loci, nonlocal/fractional diffusion, and integration of atrophy-coupled elasticity (Pal et al., 2021, Dan et al., 23 Oct 2025, Pederzoli et al., 2024).

6. Applications, Clinical Relevance, and Integration with Multimodal Data

Biophysical tau propagation models serve as a quantitative bridge from molecular misfolding to clinical phenotype:

• Personalized forecasting: Integration with tau-PET, DTI, and atrophy/morphometry enables the prediction of regional spread, atrophy, and clinical transition risk (e.g., HRS/LRS trajectories in network models) (Zhang et al., 2020, Zhang et al., 2022).
• Mechanistic insight into staging: Association with Braak regions, connectome hubs/critical nodes, and downstream atrophy patterns (Zhang et al., 2022, Corti, 2024).
• Therapeutic simulation: Modeling anti-aggregation therapy effects, active-transport blockade, or amplification/clearance modifications at the biophysical parameter level (Kevrekidis et al., 2020, Tora et al., 2024).
• Genetic and molecular targeting: Multi-layer models correlate network-dominant pathways with expression levels of inflammation, lysosomal, and neurodevelopmental genes (CHUK, TMEM106B, MCL1, NOTCH1, TH), and with clinical risk factors such as APOE (Dan et al., 23 Oct 2025).

Tau propagation modeling is central to the mechanistic systems neuroscience of AD, providing a hierarchy of biophysically interpretable frameworks—from minimal reaction–diffusion PDEs and two-species heterodimers to connectome-based, multi-layer, active-transport, and neural-operator surrogates—that quantitatively explain and predict the spatial–temporal evolution of tau pathology and its clinical consequences.