Neural operator-based digital twins for modeling amyloid-β and tau propagation and treatment optimization in Alzheimer’s disease

Research Question and Approach

Predicting the progression of Alzheimer’s disease (AD) is hindered by the complex, nonlinear nature of protein aggregation and the limitations of sparse, noisy longitudinal clinical data. This study addresses these challenges by constructing patient-specific digital twins—computational models that mirror an individual's disease state. The authors utilize a Laplacian Eigenfunction Neural Operator (LENO) to learn the underlying reaction-diffusion equations governing the spread of amyloid-beta and tau directly from PET imaging data. By treating the cortical surface as a manifold, the model captures how these proteins propagate spatially over time.

The proposed framework demonstrates high predictive accuracy, achieving 87% for amyloid-beta and 81% for tau. By integrating these learned dynamics into a PDE-constrained optimal control problem, the authors can simulate personalized treatment strategies. These strategies allow for the optimization of dosing schedules that balance the reduction of pathological protein burden against the potential side effects of treatment. The study also introduces an immersive virtual reality platform, which provides clinicians with an interactive interface to visualize patient-specific disease trajectories and compare the projected outcomes of various intervention scenarios.

This work bridges the gap between mechanistic mathematical modeling and data-driven machine learning in neurodegeneration. By providing a platform that is both predictive and prescriptive, the framework offers a foundation for precision medicine in Alzheimer’s disease. It moves beyond static diagnosis by allowing for the simulation of how a specific patient’s brain might respond to different therapeutic interventions, potentially aiding in the development of more effective, individualized treatment plans.