Demonstration of an open‐source toolbox for network spreading models: regional amyloid burden promotes tau production in Alzheimer's disease

Elinor Thompson, Anna Schroder, Tiantian He, Antoine Legouhy, Xin Zhao, James H. Cole, Neil P. Oxtoby, Daniel C. AlexanderAlzheimer 2024 · DOI

TL;DR

This paper introduces an open-source toolbox for modeling how neurodegenerative diseases spread through the brain. They demonstrate that incorporating regional amyloid burden significantly improves the model's ability to predict tau pathology, suggesting amyloid actively promotes tau aggregation, especially in later stages of Alzheimer's disease.

What Problem Does This Paper Address?

Neurodegenerative diseases like Alzheimer's are characterized by the spread of pathological proteins (like amyloid and tau) through the brain's neural network. Understanding the mechanisms driving this spread is crucial for developing effective treatments. While connectome-based models exist to simulate this spread, it's challenging to accurately capture the complex interplay of factors influencing pathology progression. Specifically, researchers want to know if local amyloid accumulation directly influences the production or spread of tau pathology. This paper addresses this by developing and validating a flexible modeling framework to test such hypotheses.

Methods

The authors developed an open-source toolbox containing three network spreading models:

  1. Network Diffusion Model (NDM): A basic model for spread between connected brain regions.
  2. Fisher-Kolmogorov-Petrovsky-Piscounov (FKPP) Model: This model incorporates both spread and a uniform production rate of pathology.
  3. Weighted-FKPP Model: This is the novel contribution. It's an extension of the FKPP model where the production rate of tau is made regionally specific, using the local amyloid burden as a weighting factor.

They used Gaussian processes optimization to fit model parameters (like the spread rate and initial seed region) to the data. To validate the significance of the amyloid weighting, they performed a null modeling test by shuffling amyloid maps.

Datasets

The study utilizes data from the Alzheimer's Disease Neuroimaging Initiative (ADNI). Specifically, they used regional tau-PET and amyloid-PET Standardized Uptake Value Ratios (SUVRs) downloaded in September 2023. Subjects were categorized into three groups based on their amyloid and tau status: amyloid-negative/tau-negative, amyloid-positive/tau-negative, and amyloid-positive/tau-positive.

Key Insights and Ideas

  • Amyloid as a Driver of Tau Production: The most significant finding is that regional amyloid burden acts as a crucial factor in promoting tau production. This moves beyond simply viewing amyloid as a precursor or co-pathology to suggesting it actively fuels tau aggregation.
  • Model Flexibility is Key: The development of a toolbox that allows comparison of different spreading models and parameters is essential for advancing our understanding of complex biological processes.
  • Stage-Dependent Effects: The impact of amyloid on tau production appears to be more pronounced in later stages of the disease (i.e., when both amyloid and tau are present).
  • Open-Source Contribution: Making the toolbox publicly available on GitHub is a valuable contribution to the research community, enabling reproducibility and further development.

Key Results

  • The weighted-FKPP model, which uses regional amyloid burden to modulate tau production, significantly improved model fit compared to the standard FKPP and NDM models across all disease stages.
  • The reduction in squared model residuals with the weighted FKPP model was statistically significant, particularly in the amyloid-positive, tau-positive group.
  • For all groups, the weighted-FKPP model showed significantly improved model fit (measured by Pearson's R) compared to the null model, with p-values less than 1x10⁻¹⁰.
  • The AICc weights (a measure of model fit considering complexity) strongly favored the amyloid-weighted FKPP model, assigning it a weight of 1.0 across all disease stages.

Limitations and Future Work

  • Model Simplification: The models are still simplifications of complex biological processes. The "production weight" is directly tied to amyloid, and other factors influencing tau production are not explicitly modeled.
  • Cross-Sectional Data: The study uses cross-sectional ADNI data. While it models spread, longitudinal data would provide a more direct validation of the dynamic progression.
  • Future Work: The authors suggest that this framework can be extended to investigate other pathological cascades and to test the efficacy of potential interventions by simulating their impact on disease spread. Further refinement of the production weighting mechanism could also be explored.

Summary updated

Abstract

Abstract Background Connectome‐based models of disease propagation are used to probe mechanisms of pathology spread in neurodegenerative disease. We present our network spreading model toolbox that allows the user to compare model fits across different models and parameters. We apply the toolbox to assess whether local amyloid levels affect production of pathological tau. Methods Data: Regional tau‐PET and amyloid‐PET SUVRs from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) study were downloaded in September 2023 (demographics in Figure 1). Subjects were separated into groups based on AT disease stage: 1) amyloid‐negative, tau‐negative; 2) amyloid‐positive, tau‐negative; 3) amyloid‐positive, tau‐positive. Positivity thresholds were defined using Gaussian mixture modelling. Data were averaged within groups to produce stage‐specific maps. Spreading models: We compared three models: the network diffusion model, Fisher‐Kolmogorov‐Petrovsky‐Piscounov (FKPP) and weighted‐FKPP models (Figure 2). Previously we showed that weighting the production rate of FKPP with regional vectors improves the model fit by accounting for different regional characteristics (He et al, MICCAI 2023). Here, we demonstrate this new “weighted‐FKPP” model, using regional amyloid as the fixed production weight. Models were initialised with unity pathology in a bilateral seed region, zero elsewhere. Parameter optimisation: Free parameters were fitted using Gaussian processes optimisation (α and seed region). NDM toolbox: All models are available in our toolbox: https://github.com/ethompson93/network_spreading_models . Null Modelling: we repeated the weighted‐FKPP model optimisation with a distribution of 100 randomly shuffled amyloid maps to demonstrate the spatial significance of the amyloid weighting. Result Adding amyloid weighting to the production term in the FKPP model improved the model fit to the data for each group. The reduction in squared model residuals with the weighted FKPP model was significant only in the amyloid‐positive tau‐positive group (Figure 3). For all groups, the results from the FKPP weighted with amyloid had significantly improved model fit (Pearson’s R) compared to the null model (all p‐values<1×10‐10). Conclusion Using our toolbox, we found that incorporating regional amyloid levels improves pathology‐spreading model fit to measured tau‐PET data, particularly at later stages of AD. This supports the hypothesis that amyloid promotes tau aggregation.
BibTeX
@inproceedings{thompson2024demonstration,
  author = {Elinor Thompson and Anna Schroder and Tiantian He and Antoine Legouhy and Xin Zhao and James H. Cole and Neil P. Oxtoby and Daniel C. Alexander},
  title = {{Demonstration of an open‐source toolbox for network spreading models: regional amyloid burden promotes tau production in Alzheimer's disease}},
  booktitle = {Alzheimer},
  year = {2024},
  doi = {10.1002/alz.093791},
}
Figure