Alzheimer’s disease is a neurodegenerative disorder characterized by the presence of amyloid-𝛽
plaques and the accumulation of misfolded tau proteins and neurofibrillary tangles in the brain. A
thorough understanding of the local accumulation of tau is critical to develop effective therapeutic
strategies. Tau pathology has traditionally been described using reaction-diffusion models, which
succeed in capturing the global spread, but fail to accurately describe the local aggregation dynamics.
Current mathematical models enforce a single-peak behavior in tau aggregation, which does not align
well with clinical observations. Here we identify a more accurate description of tau aggregation
that reflects the complex patterns observed in patients. We propose an innovative approach that
uses constitutive neural networks to autonomously discover bell-shaped aggregation functions with
multiple peaks from clinical positron emission tomography (PET) data of misfolded tau protein.
Our method reveals previously overlooked two-stage aggregation dynamics by uncovering a twoterm ordinary differential equation that links the local accumulation rate to the tau concentration.
When trained on data from amyloid-𝛽 positive and negative subjects, the neural network clearly
distinguishes between both groups and uncovers a more subtle relationship between amyloid-𝛽 and
tau than previously postulated. In line with the amyloid-tau dual pathway hypothesis, our results show
that the presence of toxic amyloid-𝛽 influences the accumulation of tau, particularly in the earlier
disease stages. We expect that our approach to autonomously discover the accumulation dynamics of
pathological proteins will improve simulations of tau dynamics in Alzheimer’s disease and provide
new insights into disease progression.
automated model discovery
,alzheimer’s disease
,neurodegeneration
,tau pathology
,constitutive neural networks
